Uploaded by bffaf01

Data Governance

advertisement
Data Governance
The Definitive Guide
People, Processes, and Tools to Operationalize
Data Trustworthiness
Evren Eryurek, Uri Gilad,
Valliappa Lakshmanan,
Anita Kibunguchy-Grant
& Jessi Ashdown
Praise for Data Governance: The Definitive Guide
We live in a digital world. Whether we realize it or not, we are living in the throes of one
of the biggest economic and social revolutions since the Industrial Revolution. It’s about
transforming traditional business processes—many of them previously nondigital or
manual—into processes that will fundamentally change how we live, how we operate our
businesses, and how we deliver value to our customers. Data defines, informs, and
predicts—it is used to penetrate new markets, control costs, drive revenues, manage risk,
and help us discover the world around us. But to realize these benefits, data must be
properly managed and stewarded. Data Governance: The Definitive Guide walks you
through the many facets of data management and data governance—people, process and
tools, data ownership, data quality, data protection, privacy and security—and does it in a
way that is practical and easy to follow. A must-read for the data professional!
—John Bottega, president of the EDM Council
Enterprises are increasingly evolving as insight-driven businesses, putting pressure on
data to satisfy new use cases and business ecosystems. Add to this business complexity,
market disruption, and demand for speed, and data governance is front and center to
make data trusted, secure, and relevant. This is not your grandfather’s slow and
bureaucratic data governance either. This book shares the secrets into how modern data
governance ensures data is the cornerstone to your business resilience, elasticity, speed,
and growth opportunity and not an afterthought.
—Michele Goetz, vice president/principal analyst–business
insights at Forrester
Data governance has evolved from a discipline focused on cost and compliance to one
that propels organizations to grow and innovate. Today’s data governance solutions can
benefit from technological advances that establish a continuous, autonomous, and
virtuous cycle. This in turn becomes an ecosystem—a community in which data is used
for good, and doing the right thing is also the easy thing. Executives looking to use data as
an asset and deliver positive business outcomes need to rethink governance’s role and
adopt the modern and transformative approach Data Governance: The Definitive Guide
provides.
—Jim Cushman, CPO of Collibra
Data Governance:
The Definitive Guide
People, Processes, and Tools to Operationalize
Data Trustworthiness
Evren Eryurek, Uri Gilad, Valliappa Lakshmanan,
Anita Kibunguchy-Grant, and Jessi Ashdown
Beijing
Boston Farnham Sebastopol
Tokyo
Data Governance: The Definitive Guide
by Evren Eryurek, Uri Gilad, Valliappa Lakshmanan, Anita Kibunguchy-Grant, and Jessi Ashdown
Copyright © 2021 Uri Gilad, Jessi Ashdown, Valliappa Lakshmanan, Evren Eryurek, and Anita
Kibunguchy-Grant. All rights reserved.
Printed in the United States of America.
Published by O’Reilly Media, Inc., 1005 Gravenstein Highway North, Sebastopol, CA 95472.
O’Reilly books may be purchased for educational, business, or sales promotional use. Online editions are
also available for most titles (http://oreilly.com). For more information, contact our corporate/institutional
sales department: 800-998-9938 or corporate@oreilly.com.
Acquisitions Editor: Jessica Haberman
Development Editor: Gary O’Brien
Production Editor: Kate Galloway
Copyeditor: Piper Editorial Consulting, LLC
Proofreader: Arthur Johnson
Indexer: WordCo Indexing Services, Inc.
Interior Designer: David Futato
Cover Designer: Karen Montgomery
Illustrator: Kate Dullea
First Edition
March 2021:
Revision History for the First Edition
2021-03-08:
First Release
See http://oreilly.com/catalog/errata.csp?isbn=9781492063490 for release details.
The O’Reilly logo is a registered trademark of O’Reilly Media, Inc. Data Governance: The Definitive Guide,
the cover image, and related trade dress are trademarks of O’Reilly Media, Inc.
The views expressed in this work are those of the authors, and do not represent the publisher’s views.
While the publisher and the authors have used good faith efforts to ensure that the information and
instructions contained in this work are accurate, the publisher and the authors disclaim all responsibility
for errors or omissions, including without limitation responsibility for damages resulting from the use of
or reliance on this work. Use of the information and instructions contained in this work is at your own
risk. If any code samples or other technology this work contains or describes is subject to open source
licenses or the intellectual property rights of others, it is your responsibility to ensure that your use
thereof complies with such licenses and/or rights.
978-1-492-06349-0
[LSI]
Table of Contents
Preface. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi
1. What Is Data Governance?. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
What Data Governance Involves
Holistic Approach to Data Governance
Enhancing Trust in Data
Classification and Access Control
Data Governance Versus Data Enablement and Data Security
Why Data Governance Is Becoming More Important
The Size of Data Is Growing
The Number of People Working and/or Viewing the Data Has Grown
Exponentially
Methods of Data Collection Have Advanced
More Kinds of Data (Including More Sensitive Data) Are Now
Being Collected
The Use Cases for Data Have Expanded
New Regulations and Laws Around the Treatment of Data
Ethical Concerns Around the Use of Data
Examples of Data Governance in Action
Managing Discoverability, Security, and Accountability
Improving Data Quality
The Business Value of Data Governance
Fostering Innovation
The Tension Between Data Governance and Democratizing Data Analysis
Manage Risk (Theft, Misuse, Data Corruption)
Regulatory Compliance
Considerations for Organizations as They Think About Data Governance
Why Data Governance Is Easier in the Public Cloud
2
4
5
6
8
9
9
10
10
13
14
16
16
17
18
19
23
23
24
25
26
28
30
v
Location
Reduced Surface Area
Ephemeral Compute
Serverless and Powerful
Labeled Resources
Security in a Hybrid World
Summary
31
32
32
32
33
34
34
2. Ingredients of Data Governance: Tools. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
The Enterprise Dictionary
Data Classes
Data Classes and Policies
Data Classification and Organization
Data Cataloging and Metadata Management
Data Assessment and Profiling
Data Quality
Lineage Tracking
Key Management and Encryption
Data Retention and Data Deletion
Workflow Management for Data Acquisition
IAM—Identity and Access Management
User Authorization and Access Management
Summary
37
38
40
43
44
45
46
46
47
50
52
52
54
55
3. Ingredients of Data Governance: People and Processes. . . . . . . . . . . . . . . . . . . . . . . . . . . 57
The People: Roles, Responsibilities, and Hats
User Hats Defined
Data Enrichment and Its Importance
The Process: Diverse Companies, Diverse Needs and Approaches to Data
Governance
Legacy
Cloud Native/Digital Only
Retail
Highly Regulated
Small Companies
Large Companies
People and Process Together: Considerations, Issues, and Some Successful
Strategies
Considerations and Issues
Processes and Strategies with Varying Success
Summary
vi
|
Table of Contents
57
58
65
65
66
67
67
69
71
72
73
74
77
84
4. Data Governance over a Data Life Cycle. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
What Is a Data Life Cycle?
Phases of a Data Life Cycle
Data Creation
Data Processing
Data Storage
Data Usage
Data Archiving
Data Destruction
Data Life Cycle Management
Data Management Plan
Applying Governance over the Data Life Cycle
Data Governance Framework
Data Governance in Practice
Example of How Data Moves Through a Data Platform
Operationalizing Data Governance
What Is a Data Governance Policy?
Importance of a Data Governance Policy
Developing a Data Governance Policy
Data Governance Policy Structure
Roles and Responsibilities
Step-by-Step Guidance
Considerations for Governance Across a Data Life Cycle
Summary
85
86
87
88
88
88
89
89
90
90
92
92
94
97
100
101
102
103
103
105
106
108
111
5. Improving Data Quality. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
What Is Data Quality?
Why Is Data Quality Important?
Data Quality in Big Data Analytics
Data Quality in AI/ML Models
Why Is Data Quality a Part of a Data Governance Program?
Techniques for Data Quality
Scorecard
Prioritization
Annotation
Profiling
Summary
113
114
116
117
121
121
123
123
123
124
131
6. Governance of Data in Flight. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
Data Transformations
Lineage
Why Lineage Is Useful
133
134
135
Table of Contents
|
vii
How to Collect Lineage
Types of Lineage
The Fourth Dimension
How to Govern Data in Flight
Policy Management, Simulation, Monitoring, Change Management
Audit, Compliance
Summary
135
136
138
139
141
141
142
7. Data Protection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
Planning Protection
Lineage and Quality
Level of Protection
Classification
Data Protection in the Cloud
Multi-Tenancy
Security Surface
Virtual Machine Security
Physical Security
Network Security
Security in Transit
Data Exfiltration
Virtual Private Cloud Service Controls (VPC-SC)
Secure Code
Zero-Trust Model
Identity and Access Management
Authentication
Authorization
Policies
Data Loss Prevention
Encryption
Differential Privacy
Access Transparency
Keeping Data Protection Agile
Security Health Analytics
Data Lineage
Event Threat Detection
Data Protection Best Practices
Separated Network Designs
Physical Security
Portable Device Encryption and Policy
Data Deletion Process
Summary
viii
| Table of Contents
143
144
145
146
146
147
147
148
149
151
151
153
155
156
157
158
158
159
160
161
162
164
165
165
165
166
167
167
168
168
170
170
173
8. Monitoring. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
What Is Monitoring?
Why Perform Monitoring?
What Should You Monitor?
Data Quality Monitoring
Data Lineage Monitoring
Compliance Monitoring
Program Performance Monitoring
Security Monitoring
What Is a Monitoring System?
Analysis in Real Time
System Alerts
Notifications
Reporting/Analytics
Graphic Visualization
Customization
Monitoring Criteria
Important Reminders for Monitoring
Summary
175
176
179
179
180
182
183
185
187
187
187
187
188
188
188
189
190
191
9. Building a Culture of Data Privacy and Security. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
Data Culture: What It Is and Why It’s Important
Starting at the Top—Benefits of Data Governance to the Business
Analytics and the Bottom Line
Company Persona and Perception
Intention, Training, and Communications
A Data Culture Needs to Be Intentional
Training: Who Needs to Know What
Beyond Data Literacy
Motivation and Its Cascading Effects
Maintaining Agility
Requirements, Regulations, and Compliance
The Importance of Data Structure
Scaling the Governance Process Up and Down
Interplay with Legal and Security
Staying on Top of Regulations
Communication
Interplay in Action
Agility Is Still Key
Incident Handling
When “Everyone” Is Responsible, No One Is Responsible
Importance of Transparency
Table of Contents
193
194
195
195
196
197
197
200
200
201
202
202
203
203
204
204
204
205
205
205
206
|
ix
What It Means to Be Transparent
Building Internal Trust
Building External Trust
Setting an Example
Summary
206
206
207
208
208
A. Google’s Internal Data Governance. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
B. Additional Resources. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
Index. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
x
|
Table of Contents
Preface
In recent years, the ease of moving to the cloud has motivated and energized a fastgrowing community of data consumers to collect, capture, store, and analyze data for
insights and decision making. For a number of reasons, as adoption of cloud comput‐
ing continues to grow, information management stakeholders have questions about
the potential risks involved in managing their data in the cloud. Evren faced such
questions for the first time when he worked in healthcare and had to put in place the
processes and technologies to govern data. Now at Google Cloud, Uri and Lak also
answer these questions nearly every week and dispense advice on getting value from
data, breaking down data silos, preserving anonymity, protecting sensitive informa‐
tion, and improving the trustworthiness of data.
We noticed that GDPR was what precipitated a sea change in customers’ behavior.
Some customers even deleted their data, thinking it was the right thing to do. That
reaction, more than any other, prompted us to write this book capturing the advice
we have provided over the years to Google Cloud customers. If data is the new cur‐
rency, we do not want enterprises to be scared of it. If the data is locked away or is not
trustworthy, it is of no value.
We all pride ourselves on helping Google Cloud customers get value for their techni‐
cal expenditures. Data is a huge investment, and we felt obligated to provide our cus‐
tomers with the best way to get value from it.
Customers’ questions usually involve one of three risk factors:
Securing the data
Storing data in a public cloud infrastructure might concern large enterprises that
typically deploy their systems on-premises and expect tight security. With a sig‐
nificant number of security threats and breaches in the news, organizations are
concerned that they might be the next victim. These factors contribute to risk
management concerns for protecting against unauthorized access to or exposure
of sensitive data, ranging from personally identifiable information (PII) to corpo‐
rate confidential information, trade secrets, or intellectual property.
xi
Regulations and compliance
There is a growing set of regulations, including the California Consumer Privacy
Act (CCPA), the European Union’s General Data Protection Regulation (GDPR),
and industry-specific standards such as global Legal Entity Identifier (LEI) num‐
bers in the financial industry and ACORD data standards in the insurance indus‐
try. Compliance teams responsible for adhering to these regulations and
standards may have concerns about oversight and control of data stored in the
cloud.
Visibility and control
Data management professionals and data consumers sometimes lack visibility
into their own data landscape: which data assets are available, where those assets
are located and how and if they can be used, and who has access to the data and
whether they should have access to it. This uncertainty limits their ability to fur‐
ther leverage their own data to improve productivity or drive business value.
These risk factors clearly highlight the need for increased data assessment, cataloging
of metadata, access control management, data quality, and information security as
core data governance competencies that the cloud provider should not only provide
but also continuously upgrade in a transparent way. In essence, addressing these risks
without abandoning the benefits provided by cloud computing has elevated the
importance of not only understanding data governance in the cloud, but also know‐
ing what is important. Good data governance can inspire customer trust and lead to
vast improvements in customer experience.
Why Your Business Needs Data Governance in the Cloud
As your business generates more data and moves it into the cloud, the dynamics of
data management change in a number of fundamental ways. Organizations should
take note of the following:
Risk management
There are concerns about potential exposure of sensitive information to unau‐
thorized individuals or systems, security breaches, or known personnel accessing
data under the wrong circumstances. Organizations are looking to minimize this
risk, so additional forms of protection (such as encryption) to obfuscate the data
object’s embedded information are required to safeguard the data should a sys‐
tem breach occur. In addition, other tools are required in order to support access
management, identify sensitive data assets, and create a policy around their
protection.
Data proliferation
The speed at which businesses create, update, and stream their data assets has
increased, and while cloud-based platforms are capable of handling increased
xii
|
Preface
data velocity, volume, and variety, it is important to introduce controls and
mechanisms to rapidly validate the quality aspects of high-bandwidth data
streams.
Data management
The need to adopt externally produced data sources and data streams (including
paid feeds from third parties) means that you should be prepared not to trust all
external data sources. You may need to introduce tools that document data line‐
age, classification, and metadata to help your employees (data consumers, in par‐
ticular) to determine data usability based on their knowledge of how the data
assets were produced.
Discovery (and data awareness)
Moving data into any kind of data lake (cloud-based or on-premises) runs the
risk of losing track of which data assets have been moved, the characteristics of
their content, and details about their metadata. The ability, therefore, to assess
data asset content and sensitivity (no matter where the data is) becomes very
important.
Privacy and compliance
Regulatory compliance demands auditable and measurable standards and proce‐
dures that ensure compliance with internal data policies as well as external gov‐
ernment regulations. Migrating data to the cloud means that organizations need
tools to enforce, monitor, and report compliance, as well as ensure that the right
people and services have access and permissions to the right data.
Framework and Best Practices for Data Governance
in the Cloud
Given the changing dynamics of data management, how should organizations think
about data governance in the cloud, and why is it important? According to
TechTarget, data governance is
the overall management of the availability, usability, integrity, and security of data used
in an enterprise. A sound data governance program includes a governing body or
council, a defined set of procedures and a plan to execute those procedures.1
Simply put, data governance encompasses the ways that people, processes and tech‐
nology can work together to enable auditable compliance with defined and agreedupon data policies.
1 Craig Stedman and Jack Vaughan, “What Is Data Governance and Why Does It Matter?” TechTarget, Decem‐
ber 2019. This article was updated in February 2020; the current version no longer includes this quote.
Preface
|
xiii
Data Governance Framework
Enterprises need to think about data governance comprehensively, from data intake
and ingestion to cataloging, persistence, retention, storage management, sharing,
archiving, backup, recovery, loss prevention, disposition, and removal and deletion:
Data discovery and assessment
Cloud-based environments often offer an economical option for creating and
managing data lakes, but the risk remains for ungoverned migration of data
assets. This risk represents a potential loss of knowledge of what data assets are in
the data lake, what information is contained within each object, and where those
data objects originated from. A best practice for data governance in the cloud is
data discovery and assessment in order to know what data assets you have. The
data discovery and assessment process is used to identify data assets within the
cloud environment, and to trace and record each data asset’s origin and lineage,
what transformations have been applied, and object metadata. (Often this meta‐
data describes the demographic details, such as the name of the creator, the size
of the object, the number of records if it is a structured data object, or when it
was last updated.)
Data classification and organization
Properly evaluating a data asset and scanning the content of its different
attributes can help categorize the data asset for subsequent organization. This
process can also infer whether the object contains sensitive data and, if so, clas‐
sify it in terms of the level of data sensitivity, such as personal and private data,
confidential data, or intellectual property. To implement data governance in the
cloud, you’ll need to profile and classify sensitive data to determine which gover‐
nance policies and procedures apply to the data.
Data cataloging and metadata management
Once your data assets are assessed and classified, it is crucial that you document
your learnings so that your communities of data consumers have visibility into
your organization’s data landscape. You need to maintain a data catalog that con‐
tains structural metadata, data object metadata, and the assessment of levels of
sensitivity in relation to the governance directives (such as compliance with one
or more data privacy regulations). The data catalog not only allows data consum‐
ers to view this information but can also serve as part of a reverse index for
search and discovery, both by phrase and (given the right ontologies) by concept.
It is also important to understand the format of structured and semi-structured
data objects and allow your systems to handle these data types differently, as
necessary.
xiv
|
Preface
Data quality management
Different data consumers may have different data quality requirements, so it’s
important to provide a means of documenting data quality expectations as well as
techniques and tools for supporting the data validation and monitoring process.
Data quality management processes include creating controls for validation,
enabling quality monitoring and reporting, supporting the triage process for
assessing the level of incident severity, enabling root cause analysis and recom‐
mendation of remedies to data issues, and data incident tracking. The right pro‐
cesses for data quality management will provide measurably trustworthy data for
analysis.
Data access management
There are two aspects of governance for data access. The first aspect is the provi‐
sioning of access to available assets. It’s important to provide data services that
allow data consumers to access their data, and fortunately, most cloud platforms
provide methods for developing data services. The second aspect is prevention of
improper or unauthorized access. It’s important to define identities, groups, and
roles and assign access rights to establish a level of managed access. This best
practice involves managing access services as well as interoperating with the
cloud provider’s identity and access management (IAM) services by defining
roles, specifying access rights, and managing and allocating access keys to ensure
that only authorized and authenticated individuals and systems are able to access
data assets according to defined rules.
Auditing
Organizations must be able to assess their systems to make sure that they are
working as designed. Monitoring, auditing, and tracking (who did what and
when and with what information) helps security teams gather data, identify
threats, and act on those threats before they result in business damage or loss. It’s
important to perform regular audits to check the effectiveness of controls in
order to quickly mitigate threats and evaluate overall security health.
Data protection
Despite the efforts of information technology security groups to establish perim‐
eter security as a way to prevent unauthorized individuals from accessing data,
perimeter security is not and never has been sufficient for protecting sensitive
data. While you might be successful in preventing someone from breaking into
your system, you are not protected from an insider security breach or even from
exfiltration (data theft). It’s important to institute additional methods of data pro‐
tection—including encryption at rest, encryption in transit, data masking, and
permanent deletion—to ensure that exposed data cannot be read.
Preface
|
xv
Operationalizing Data Governance in Your Organization
Technology certainly helps support the data governance principles presented in the
preceding section, but data governance goes beyond the selection and implementa‐
tion of products and tools. The success of a data governance program depends on a
combination of:
• People to build the business case, develop the operating model, and take on
appropriate roles
• Processes that operationalize policy development, implementation, and
enforcement
• Technology used to facilitate the ways that people execute those processes
The following steps are critical in planning, launching, and supporting a data gover‐
nance program:
1. Build the business case. Establish the business case by identifying critical busi‐
ness drivers to justify the effort and investment associated with data governance.
Outline perceived data risks (such as the storage of data on cloud-based plat‐
forms) and indicate how data governance helps the organization mitigate those
risks.
2. Document guiding principles. Assert core principles associated with governance
and oversight of enterprise data. Document those principles in a data governance
charter to present to senior management.
3. Get management buy-in. Engage data governance champions and get buy-in
from the key senior stakeholders. Present your business case and guiding princi‐
ples to C-level management for approval.
4. Develop an operating model. Once you have management approval, define the
data governance roles and responsibilities, and then describe the processes and
procedures for the data governance council and data stewardship teams who will
define processes for defining and implementing policies as well as reviewing and
remediating identified data issues.
5. Establish a framework for accountability. Establish a framework for assigning
custodianship and responsibility for critical data domains. Make sure there is vis‐
ibility to the “data owners” across the data landscape. Provide a methodology to
ensure that everyone is accountable for contributing to data usability.
6. Develop taxonomies and ontologies. There may be a number of governance
directives associated with data classification, organization, and—in the case of
sensitive information—data protection. To enable your data consumers to com‐
ply with those directives, there must be a clear definition of the categories (for
organizational structure) and classifications (for assessing data sensitivity).
xvi
|
Preface
7. Assemble the right technology stack. Once you’ve assigned data governance
roles to your staff and defined and approved your processes and procedures, you
should assemble a suite of tools that facilitate ongoing validation of compliance
with data policies and accurate compliance reporting.
8. Establish education and training. Raise awareness of the value of data gover‐
nance by developing educational materials highlighting data governance practi‐
ces and procedures, and the use of supporting technology. Plan for regular
training sessions to reinforce good data governance practices.
The Business Benefits of Robust Data Governance
Data security, data protection, data accessibility and usability, data quality, and other
aspects of data governance will continue to emerge and grow as critical priorities for
organizations. And as more organizations migrate their data assets to the cloud, the
need for auditable practices for ensuring data utility will also continue to grow. To
address these directives, businesses should frame their data governance practices
around three key components:
• A framework that enables people to define, agree to, and enforce data policies
• Effective processes for control, oversight, and stewardship over all data assets
across on-premises systems, cloud storage, and data warehouse platforms
• The right tools and technologies for operationalizing data policy compliance
With this framework in mind, an effective data governance strategy and operating
model provides a path for organizations to establish control and maintain visibility
into their data assets, providing a competitive advantage over their peers. Organiza‐
tions will likely reap immense benefits as they promote a data-driven culture within
their organizations—specifically:
Improved decision making
Better data discovery means that users can find the data they need when they
need it, which makes them more efficient. Data-driven decision making plays a
huge role in improving business planning within an organization.
Better risk management
A good data governance operating model helps organizations audit their pro‐
cesses more easily so that they reduce the risk of fines, increase customer trust,
and improve operations. Downtime can be minimized while productivity still
grows.
Preface
|
xvii
Regulatory compliance
Increasing governmental regulation has made it even more important for organi‐
zations to establish data governance practices. With a good data governance
framework, organizations can embrace the changing regulatory environment
instead of simply reacting to it.
As you migrate more of your data to the cloud, data governance provides a level of
protection against data misuse. At the same time, auditable compliance with defined
data policies helps demonstrate to your customers that you protect their private
information, alleviating their concerns about information risks.
Who Is This Book For?
The current growth in data is unprecedented and, when coupled with increased regu‐
lations and fines, has meant that organizations are forced to look into their data gov‐
ernance plans to make sure that they do not become the next statistic. Therefore,
every organization will need to establish an understanding of the data it collects, the
liability and regulation associated with that data, and who has access to it. This book
is for you if you want to know what that entails, the risks to be aware of, and the con‐
siderations to keep in mind.
This book is for anyone who needs to implement the processes or technology that
enables data to become trustworthy. This book covers the ways that people, processes,
and technology can work together to enable auditable compliance with defined and
agreed-upon data policies.
The benefits of data governance are multifaceted, ranging from legal and regulatory
compliance to better risk management and the ability to drive top-line revenue and
cost savings by creating new products and services. Read this book to learn how to
establish control and maintain visibility into your data assets, which will provide you
with a competitive advantage over your peers.
Conventions Used in This Book
The following typographical conventions are used in this book:
Italic
Indicates new terms, URLs, email addresses, filenames, and file extensions.
Constant width
Used for program listings, as well as within paragraphs to refer to program ele‐
ments such as variable or function names, databases, data types, environment
variables, statements, and keywords.
xviii
|
Preface
This element signifies a tip or suggestion.
This element signifies a general note.
This element indicates a warning or caution.
O’Reilly Online Learning
For more than 40 years, O’Reilly Media has provided technol‐
ogy and business training, knowledge, and insight to help
companies succeed.
Our unique network of experts and innovators share their knowledge and expertise
through books, articles, and our online learning platform. O’Reilly’s online learning
platform gives you on-demand access to live training courses, in-depth learning
paths, interactive coding environments, and a vast collection of text and video from
O’Reilly and 200+ other publishers. For more information, visit http://oreilly.com.
How to Contact Us
Please address comments and questions concerning this book to the publisher:
O’Reilly Media, Inc.
1005 Gravenstein Highway North
Sebastopol, CA 95472
800-998-9938 (in the United States or Canada)
707-829-0515 (international or local)
707-829-0104 (fax)
We have a web page for this book, where we list errata, examples, and any additional
information. You can access this page at https://oreil.ly/data-governance-TDG.
Preface
|
xix
Email bookquestions@oreilly.com to comment or ask technical questions about this
book.
For news and information about our books and courses, visit http://oreilly.com.
Find us on Facebook: http://facebook.com/oreilly
Follow us on Twitter: http://twitter.com/oreillymedia
Watch us on YouTube: http://www.youtube.com/oreillymedia
Acknowledgments
Thank you to our respective families, teammates, and managers. Gary O’Brien, our
editor at O’Reilly, was a force of nature—this book would not exist without his con‐
stant prodding and invaluable advice. Thanks also to our technical reviewers for their
invaluable suggestions.
xx
|
Preface
CHAPTER 1
What Is Data Governance?
Data governance is, first and foremost, a data management function to ensure the
quality, integrity, security, and usability of the data collected by an organization. Data
governance needs to be in place from the time a factoid of data is collected or gener‐
ated until the point in time at which that data is destroyed or archived. Along the way
in this full life cycle of the data, data governance focuses on making the data available
to all stakeholders in a form that they can readily access. In addition, it must be one
they can use in a manner that generates the desired business outcomes (insights, anal‐
ysis) and conforms to regulatory standards, if relevant. These regulatory standards
are often an intersection of industry (e.g., healthcare), government (e.g., privacy), and
company (e.g., nonpartisan) rules and codes of behavior. Moreover, data governance
needs to ensure that the stakeholders get a high-quality integrated view of all the data
within the enterprise. There are many facets to high-quality data—the data needs to
be correct, up to date, and consistent. Finally, data governance needs to be in place to
ensure that the data is secure, by which we mean that:
• It is accessed only by permitted users in permitted ways
• It is auditable, meaning all accesses, including changes, are logged
• It is compliant with regulations
The purpose of data governance is to enhance trust in the data. Trustworthy data is
necessary to enable users to employ enterprise data to support decision making, risk
assessment, and management using key performance indicators (KPIs). Using data,
you can increase confidence in the decision-making process by showing supporting
evidence. The principles of data governance are the same regardless of the size of the
enterprise or the quantity of data. However, data governance practitioners will make
choices with regard to tools and implementation based on practical considerations
driven by the environment within which they operate.
1
What Data Governance Involves
The advent of big data analytics, powered by the ease of moving to the cloud and the
ever-increasing capability and capacity of compute power, has motivated and ener‐
gized a fast-growing community of data consumers to collect, store, and analyze data
for insights and decision making. Nearly every computer application these days is
informed by business data. It is not surprising, therefore, that new ideas inevitably
involve the analysis of existing data in new ways, as well as the collection of new data‐
sets, whether through new systems or by purchase from external vendors. Does your
organization have a mechanism to vet new data analysis techniques and ensure that
any data collected is stored securely, that the data collected is of high quality, and that
the resulting capabilities accrue to your brand value? While it’s tempting to look only
toward the future power and possibilities of data collection and big data analytics,
data governance is a very real and very important consideration that cannot be
ignored. In 2017, Harvard Business Review reported that more than 70% of employees
have access to data they should not.1 This is not to say that companies should adopt a
defensive posture; it’s only to illustrate the importance of governance to prevent data
breaches and improper use of the data. Well-governed data can lead to measurable
benefits for an organization.
Spotify Creates Discover Weekly
As an example of how well-governed data can lead to measurable benefits to an orga‐
nization and how the availability of data can completely change an entire industry,
consider the Spotify Discover Weekly feature. In the early 2010s, the way most people
listened to music was to purchase physical/digital albums and rip them to create cus‐
tom playlists. These playlists, consisting of songs you owned, was what you listened
to.
There was also a large and thriving illegal music-sharing ecosystem that consisted of
pirated singles that could be added to your playlists. In an effort to get the pirated
music system under control, music labels allowed the sales of digital singles. As the
size of people’s digital libraries grew, and as internet connections became more relia‐
ble, consumers became willing to keep their purchased tracks online and stream them
to their audio devices. Music labels were also willing to “rent” out music when it was
streamed. Instead of selling the song, the music labels would be paid each time the
song was played.
1 Leandro DalleMule and Thomas H. Davenport, “What’s Your Data Strategy?” Harvard Business Review (May–
June 2017): 112–121.
2
|
Chapter 1: What Is Data Governance?
This was how Spotify (which is now the world’s largest music streaming service) got
started. It’s worth noting that Spotify owes its very existence to data governance. It got
started as a way for music labels to get paid for their work—music piracy was deci‐
mating the music industry. Spotify’s entire business model was built around tracking
the songs users played and reimbursing artists for those songs. The ability to prove
that its handling of the data was trustworthy is the reason that Spotify became a viable
music service in the first place.
The fact that Spotify was keeping tabs on which songs users played meant that it had
data on what people listened to. Thus it was now possible for Spotify to recommend
new songs to its listeners. Such recommendation algorithms key off of three things:
• Find other songs by the artists you listen to, or songs in the same genre (e.g.,
1940s jazz). This is called content-based recommendation.
• Find users who like the same songs you do and recommend the songs that those
users like. This is called collaborative filtering.
• Use models that analyze the raw audio files of songs you like and recommend
songs that are similar. The raw audio captures many inherent features, such as
the beat. If you tend to like music with a fast beat and with repeating tonal
phrases, the algorithm can recommend other songs with a similar structure. This
is called similarity matching.
At that point, Edward Newett, an engineer at Spotify, had an interesting idea: instead
of recommending songs one at a time, what if Spotify created a playlist of recommen‐
dations? And so, every Monday, Spotify would recommend what they thought each
individual user would like. This was called Discover Weekly.
Discover Weekly was a huge hit—within a year after its launch, more than 40 million
people had used the service and streamed nearly five billion tracks. The deep person‐
alization had worked. The music sounded familiar but was still novel. The service
allowed music lovers to discover new titles, and new artists to find audiences, and it
gave Spotify’s customers an event to look forward to every week.
Spotify was able to use its recommendation algorithm to provide artists and music
labels with insights about fans’ preferences. It could leverage the recommendation
algorithm to expand users’ music preferences and introduce novel bands. This extra
knowledge and marketing ability has allowed Spotify to negotiate with music publish‐
ers from a position of strength.
None of this would have been possible if Spotify had not assured its users that the
information about their listening habits was being used in a responsible way to
improve their own music-listening experience. European regulators are very protec‐
tive of the privacy of EU citizens. Spotify, being based in Europe, could not have got‐
ten its recommendation systems off the ground if it had not proven that it had robust
privacy controls in place, and that it was ensuring that data scientists could devise
algorithms but not breach data that could be tied to individuals.
What Data Governance Involves
|
3
Discover Weekly illustrates how data, properly governed, can create a well-loved
brand and change the power dynamic in an entire industry. Spotify extended its rec‐
ommendations with Spotify Wrapped, where listeners everywhere get a deep dive into
their most memorable listening moments of the year. This is a great way to have peo‐
ple remember and share their most listened-to songs and artists (see Figure 1-1).
Figure 1-1. Anita Kibunguchy-Grant’s 2020 Wrapped playlist
Holistic Approach to Data Governance
Several years ago, when smartphones with GPS sensors were becoming ubiquitous,
one of the authors of this book was working on machine learning algorithms to pre‐
dict the occurence of hail. Machine learning requires labeled data—something that
was in short supply at the temporal and spatial resolution the research team needed.
Our team hit on the idea of creating a mobile application that would allow citizen sci‐
entists to report hail at their location.2 This was our first encounter with making
choices about what data to collect—until then, we had mostly been at the receiving
2 This application is the Meteorological Phenomena Identification Near the Ground (mPING) Project, devel‐
oped through a partnership between NSSL, the University of Oklahoma, and the Cooperative Institute for
Mesoscale Meteorological Studies.
4
|
Chapter 1: What Is Data Governance?
end of whatever data the National Weather Service was collecting. Considering the
rudimentary state of information security tools in an academic setting, we decided to
forego all personally identifying information and make the reporting totally anony‐
mous, even though this meant that certain types of reported information became
somewhat unreliable. Even this anonymous data brought tremendous benefits—we
started to evaluate hail algorithms at greater resolutions, and this improved the qual‐
ity of our forecasts. This new dataset allowed us to calibrate existing datasets, thus
enhancing the data quality of other datasets as well. The benefits went beyond data
quality and started to accrue toward trustworthiness—involvement of citizen scien‐
tists was novel enough that National Public Radio carried a story about the project,
emphasizing the anonymous nature of the data collection.3 The data governance lens
had allowed us to carefully think about which report data to collect, improve the
quality of enterprise data, enhance the quality of forecasts produced by the National
Weather Service, and even contribute to the overall brand of our weather enterprise.
This combination of effects—regulatory compliance, better data quality, new business
opportunities, and enhanced trustworthiness—was the result of a holistic approach to
data governance.
Fast-forward a few years, and now, at Google Cloud, we are all part of a team that
builds technology for scalable cloud data warehouses and data lakes. One of the
recurring concerns that our enterprise customers have is around what best practices
and policies they should put in place to manage the classification, discovery, availabil‐
ity, accessibility, integrity, and security of their data—data governance—and custom‐
ers approach it with the same sort of apprehension that our small team in academia
did.
Yet the tools and capabilities that an enterprise has at its disposal to carry out data
governance are quite powerful and diverse. We hope to convince you that you should
not be afraid of data governance, and that properly applying data governance can
open up new worlds of possibility. While you might initially approach data gover‐
nance purely from a legal or regulatory compliance standpoint, applying governance
policies can drive growth and lower costs.
Enhancing Trust in Data
Ultimately, the purpose of data governance is to build trust in data. Data governance
is valuable to the extent that it adds to stakeholders’ trust in the data—specifically, in
how that data is collected, analyzed, published, or used.
Ensuring trust in data requires that a data governance strategy address three key
aspects: discoverability, security, and accountability (see Figure 1-2). Discoverability
3 It was on the radio, but you can read about it on NPR’s All Tech Considered blog.
What Data Governance Involves
|
5
itself requires data governance to make technical metadata, lineage information, and
a business glossary readily available. In addition, business critical data needs to be
correct and complete. Finally, master data management is necessary to guarantee that
data is finely classified and thus ensure appropriate protection against inadvertent or
malicious changes or leakage. In terms of security, regulatory compliance, manage‐
ment of sensitive data (personally identifiable information, for example), and data
security and exfiltration prevention may all be important depending on the business
domain and the dataset in question. If discoverability and security are in place, then
you can start treating the data itself as a product. At that point, accountability
becomes important, and it is necessary to provide an operating model for ownership
and accountability around boundaries of data domains.
Figure 1-2. The three key aspects of data governance that must be addressed to enhance
trust in data
Classification and Access Control
While the purpose of data governance is to increase the trustworthiness of enterprise
data so as to derive business benefits, it remains the case that the primary activity
associated with data governance involves classification and access control. Therefore,
to understand the roles involved in data governance, it is helpful to consider a typical
classification and access control setup.
Let’s take the case of protecting the human resources information of employees, as
shown in Figure 1-3.
6
|
Chapter 1: What Is Data Governance?
Figure 1-3. Protecting the human resources information of employees
The human resources information includes several data elements: each employee’s
name, their date of hire, past salary payments, the bank account into which those sal‐
ary payments were deposited, current salary, etc. Each of these data elements is pro‐
tected in different ways, depending on the classification level. Potential classification
levels might be public (things accessible by people not associated with the enterprise),
external (things accessible by partners and vendors with authorized access to the
enterprise internal systems), internal (things accessible by any employee of the orga‐
nization), and restricted. For example, information about each employee’s salary pay‐
ments and which bank account they were deposited into would be restricted to
managers in the payroll processing group only. On the other hand, the restrictions
could be more dynamic. An employee’s current salary might be visible only to their
manager, and each manager might be able to see salary information only for their
respective reports. The access control policy would specify what users can do when
they access the data—whether they can create a new record, or read, update, or delete
existing records.
The governance policy is typically specified by the group that is accountable for the
data (here, the human resources department)—this group is often referred to as the
governors. The policy itself might be implemented by the team that operates the data‐
base system or application (here, the information technology department), and so
changes such as adding users to permitted groups are often carried out by the IT
team—hence, members of that team are often referred to as approvers or data stew‐
ards. The people whose actions are being circumscribed or enabled by data gover‐
nance are often referred to as users. In businesses where not all employees have access
What Data Governance Involves
|
7
to enterprise data, the set of employees with access might be called knowledge workers
to differentiate them from those without access..
Some enterprises default to open—for example, when it comes to business data, the
domain of authorized users may involve all knowledge workers in the enterprise.
Other enterprises default to closed—business data may be available only to those with
a need to know. Policies such as these are within the purview of the data governance
board in the organization—there is no uniquely correct answer on which approach is
best.
Data Governance Versus Data Enablement and Data Security
Data governance is often conflated with data enablement and with data security.
Those topics intersect but have different emphases:
• Data governance is mostly focused on making data accessible, reachable, and
indexed for searching across the relevant constituents, usually the entire organi‐
zation’s knowledge-worker population. This is a crucial part of data governance
and will require tools such as a metadata index, a data catalog to “shop for” data.
Data governance extends data enablement into including a workflow in which
data acquisition can take place. Users can search for data by context and descrip‐
tion, find the relevant data stores, and ask for access, including the desired use
case as justification. An approver (data steward) will need to review the ask,
determine whether the ask is justified and whether the data being requested can
actually serve the use case, and kick off a process through which the data can be
made accessible.
• Data enablement goes further than making data accessible and discoverable; it
extends into tooling that allows rapid analysis and processing of the data to
answer business-related questions: “how much is the business spending on this
topic,” “can we optimize this supply chain,” and so on. The topic is crucial and
requires knowledge of how to work with data, as well as what the data actually
means—best addressed by including, from the get-go, metadata that describes the
data and includes value proposition, origin, lineage, and a contact person who
curates and owns the data in question, to allow for further inquiry.
• Data security, which intersects with both data enablement and data governance,
is normally thought about as a set of mechanics put in place to prevent and block
unauthorized access. Data governance relies on data security mechanics to be in
place but goes beyond just prevention of unauthorized access and into policies
about the data itself, its transformation according to data class (see Chapter 7),
and the ability to prove that the policies set to access and transform the data over
time are being complied with. The correct implementation of security mechanics
promotes the trust required to share data broadly or “democratize access” to the
data.
8
|
Chapter 1: What Is Data Governance?
Why Data Governance Is Becoming More Important
Data governance has been around since there was data to govern, although it was
often restricted to IT departments in regulated industries, and to security concerns
around specific datasets such as authentication credentials. Even legacy data process‐
ing systems needed a way to not only ensure data quality but also control access to
data.
Traditionally, data governance was viewed as an IT function that was performed in
silos related to data source type. For example, a company’s HR data and financial
data, typically highly controlled data sources with strictly controlled access and spe‐
cific usage guidelines, would be controlled by one IT silo, whereas sales data would be
in a different, less restrictive silo. Holistic or “centralized” data governance may have
existed within some organizations, but the majority of companies viewed data gover‐
nance as a departmental concern.
Data governance has come into prominence because of the recent introductions of
GDPR- and CCPA-type regulations that affect every industry, beyond just healthcare,
finance, and a few other regulated industries. There has also been a growing realiza‐
tion about the business value of data. Because of this, the data landscape is vastly dif‐
ferent today.
The following are just a few ways in which the topography has changed over time,
warranting very different approaches to and methods for data governance.
The Size of Data Is Growing
There is almost no limit to the kinds and amount of data that can now be collected. In
a whitepaper published in November 2018, International Data Corporation predicts
that the global datasphere will balloon to 175 ZB by 2025 (see Figure 1-4).4
This rise in data captured via technology, coupled with predictive analyses, results in
systems nearly knowing more about today’s users than the users themselves.
4 David Reinsel, John Gantz, and John Rydning, “The Digitization of the World: From Edge to Core”, Novem‐
ber 2018.
Why Data Governance Is Becoming More Important
|
9
Figure 1-4. The size of the global datasphere is expected to exhibit dramatic growth
The Number of People Working and/or Viewing the Data Has Grown
Exponentially
A report by Indeed shows that the demand for data science jobs had jumped 78%
between 2015 and 2018.5 IDC also reports that there are now over five billion people
in the world interacting with data, and it projects this number to increase to six bil‐
lion (nearly 75% of the world’s population) in 2025. Companies are obsessed with
being able to make “data-driven decisions,” requiring an inordinate amount of head‐
count: from the engineers setting up data pipelines to analysts doing data curation
and analyses, and business stakeholders viewing dashboards and reports. The more
people working and viewing data, the greater the need for complex systems to man‐
age access, treatment, and usage of data because of the greater chance of misuse of the
data.
Methods of Data Collection Have Advanced
No longer must data only be batch processed and loaded for analysis. Companies are
leveraging real-time or near real-time streaming data and analytics to provide their
customers with better, more personalized engagements. Customers now expect to
access products and services wherever they are, over whatever connection they have,
and on any device. IDC predicts that this infusion of data into business workflows
and personal streams of life will result in nearly 30% of the global datasphere to be
real-time by 2025, as shown in Figure 1-5.6
5 “The Best Jobs in the US: 2019”, Indeed, March 19, 2019.
6 Reinsel et al. “The Digitization of the World.”
10
|
Chapter 1: What Is Data Governance?
Figure 1-5. More than 25% of the global datasphere will be real-time data by 2025
The advent of streaming, however, while greatly increasing the speed to analytics, also
carries with it the potential risk of infiltration, bringing about the need for complex
setup and monitoring for protection.
Advanced Data Collection in Sports
It used to be that when you talked about sport statistics, you were talking about rela‐
tively coarse-grained data—things like wins and losses. In some sports, you might
have information about a player’s performance (for example, average number of runs
a cricketer gets per inning). However, the quantity and type of data collected in sports
has changed dramatically because teams are looking to better understand what levers
they can pull to be successful in these highly competitive fields.
It’s therefore no surprise that the National Football League (NFL) wanted to better
quantify the value of specific plays and actions within a play, which is why it started
using league-wide analytics in 2015. If you’re not familiar with American football, it is
a complex sport, primarily governed by the NFL. The NFL is a professional American
football league consisting of 32 teams that are divided equally between the National
Football Conference (NFC) and the American Football Conference (AFC).
Traditional metrics such as “yards per carry” or “total rushing yards” can be flawed;
recognizing a need to grow its analytics and data collection process, the NFL created
Next Gen Stats (NGS). NGS is a league program wherein the pads of every player and
official, along with game balls, pylons, and first down chains are all tagged with a
radio frequency identification (RFID) chip, which allows for a very robust set of
Why Data Governance Is Becoming More Important
|
11
statistics after every game. These statistics constitute real-time data for every player
on every play, in every situation, anywhere on the field, including location, speed,
acceleration, and velocity (Figure 1-6).
Figure 1-6. An example of analysis and visualization of the NFL data in Kaggle. Note‐
book by Rob Mulla
The types of questions the NFL wanted an answer to included, for example, “What
contributes to a successful run play?” It wanted to know whether success depends
mostly on the ball carrier who takes the hand off, or on their teammates (by way of
blocking), or on the coach (by way of the play call). And could the data even show
what part was played by the opposing defense and the actions it took? The NFL
would also want to predict how many yards a team might gain on given rushing plays
12
|
Chapter 1: What Is Data Governance?
as they happen. Deeper insight into rushing plays ultimately helps teams, media, and
fans better understand the skillsets of players and the strategies of coaches.
Because of this, each year the league hosts the NFL’s Big Data Bowl, a sports analytics
contest that challenges talented members of the analytics community—from college
students to professionals—to contribute to the NFL’s continuing evolution in the use
of advanced analytics.7 Contestants are asked to analyze and rethink trends and player
performance in order to innovate on the way football is played and coached.
This NFL example showcases just how much the methods of data collection have
advanced. It’s no surprise that this is what’s really accelerating the amount of data
being generated in the global datasphere.
More Kinds of Data (Including More Sensitive Data) Are Now
Being Collected
It’s projected that by 2025 every person using technology and generating data will
have more than 4,900 digital data engagements per day; that’s about about one digital
interaction every eighteen seconds (see Figure 1-7).8
Figure 1-7. By 2025, a person will interact with data-creating technology more than
4,900 times a day
Many of those interactions will include the generation and resulting collection of a
myriad of sensitive data such as social security numbers, credit card numbers, names,
addresses, and health conditions, to name a few categories. The proliferation of the
collection of these extremely sensitive types of data carries with it great customer
(and regulator) concern about how that data is used and treated, and who gets to view
it.
7 Kaggle: NFL Big Data Bowl.
8 Reinsel et al. “The Digitization of the World.”
Why Data Governance Is Becoming More Important
|
13
The Use Cases for Data Have Expanded
Companies are striving to use data to make better business decisions, coined datadriven decision making. They not only are using data internally to drive day-to-day
business execution, but are also using data to help their customers make better deci‐
sions. Amazon is an example of a company doing this via collecting and analyzing
items in customers’ past purchases, items the customers have viewed, and items in
their virtual shopping carts, as well as the items they’ve ranked/reviewed after pur‐
chase, to drive targeted messaging and recommendations for future purchases.
While this Amazon use case makes perfect business sense, there are types of data
(sensitive) coupled with specific use cases for that data that are not appropriate (or
even legal). For sensitive types of data, it matters not only how that data is treated but
also how it’s used. For example, employee data may be used/viewed internally by a
company’s HR department, but it would not be appropriate for that data to be used/
viewed by the marketing department.
Use of Data to Make Better Decisions
Many of us enjoy the suggestions our friends and family make for books, movies,
music, and so on. Have you noticed how some of these suggestions pop up while you
are searching for something online? Here we would like to highlight a few examples
that show how use of data resulted in better decisions and better business outcomes.
An example from Safari Books Online shows the use of modern analytics tools to
achieve improved sales. Safari Books Online is known for its very large and diverse
customer base and for having more than 30,000 books and videos accessible from
various platforms. Safari Books Online wanted to unlock value from its massive
amounts of usage data, user search results, and trends, and to connect all this data to
drive better sales intelligence and higher sales.
The key was to achieve this in near real-time—after all, none of us likes to wait 10
minutes for the results of an online search. In order to provide real-time insights, the
Safari Books Online team routinely transferred the usage data corresponding to the
content delivery network (CDN) to a cloud native data warehouse to make the infor‐
mation available outside the original silo (see Figure 1-8).
The Safari Books Online team wanted to drill into the data, deliver various dash‐
boards, provide a better user experience, and deliver faster ad hoc queries. Using the
new analytics made supporting users much quicker and simpler and achieved higher
customer satisfaction. This was because the team could get to relevant information
about users (such as their IP addresses, or the title of a book they were querying) in
near real-time.
14
|
Chapter 1: What Is Data Governance?
Figure 1-8. Bringing the usage data from content delivery network (CDN) and web
application logs into a smart data analytics platform to achieve better data-driven
decisions
Achieving better sales intelligence was among the most important use cases for Safari
Books Online when it began its journey into data-driven decision making. All the
data it had that was once buried, or not even available from its web logs, became sales
leads. Assessment of interest among likely readers was integrated into the CRM sys‐
tem and quickly turned into actionable information (see Figure 1-9).
Figure 1-9. A dashboard that is used to monitor titles and trends
Another good example is from California Design Den, which transformed its
decision-making process with data on pricing and inventory management. By leaning
on smart analytics platforms, they were able to achieve much faster pricing decisions,
sell their inventory, and achieve better profitability.
The ability to aggregate different types of data for decision making (while balancing
which data to retain and which to get rid of) is key. Not all data can be valuable for
better decision making. It is equally important to guard against biases when you are
trying to establish your data-driven decision-making process. You need to define your
Why Data Governance Is Becoming More Important
|
15
objectives—maybe set some easy-to-measure goals at first—and create a list of highvalue questions to which you want to elicit answers. It is OK to go back and revisit
your starting point, objectives, and metrics. The answers are to be found in your data,
but looking at it from different perspectives will help you decide which data is
relevant.
The world is your oyster when it comes to gaining insight from your data. Asking the
high-value questions to obtain deeper insights is the essential part of the value chain
when it comes to data-driven decision making. Whether you want to drive sales intel‐
ligence and increase your revenue, improve support and customer experience, or
detect malicious use to prevent operational issues, data-driven decision making is key
for any business and any operation, and there are various smart tools out there to
help you start taking advantage of your invaluable data.
New Regulations and Laws Around the Treatment of Data
The increase in data and data availability has resulted in the desire and need for regu‐
lations on data, data collection, data access, and data use. Some regulations that have
been around for quite some time—for example, the Health Insurance Portability and
Accountability Act of 1996 (HIPAA), the law protecting the collection and use of per‐
sonal health data—not only are well known, but companies that have had to comply
with them have been doing so for decades—meaning their processes and methodol‐
ogy for treatment of this sensitive data are fairly sophisticated. New regulations, such
as the EU’s General Data Protection Regulation (GDPR) and the California Con‐
sumer Privacy Act (CCPA) in the US, are just two examples of usage and collection
controls that apply to myriad companies, for many of whom such governance of data
was not baked into their original data architecture strategy. Because of this, compa‐
nies that have not had to worry about regulatory compliance before have a more dif‐
ficult time modifying their technology and business processes to maintain
compliance with these new regulations.
Ethical Concerns Around the Use of Data
While use cases themselves can fit into the category of ethical use of data, new tech‐
nology around machine learning and artificial intelligence has spawned new concerns
around the ethical use of data.
One recent example from 2018 is that of Elaine Herzberg, who, while wheeling her
bike across a street in Tempe, Arizona, was struck and killed by a self-driving car.9
This incident raised questions about responsibility. Who was responsible for Elaine’s
9 Aarian Marshall and Alex Davies, “Uber’s Self-Driving Car Saw the Woman It Killed, Report Says”, Wired,
May 24, 2018.
16
|
Chapter 1: What Is Data Governance?
death? The person in the driver’s seat? The company testing the car’s capabilities? The
designers of the AI system?
While not deadly, consider the following additional examples:
• In 2014, Amazon developed a recruiting tool for identifying software engineers it
might want to hire; however, it was found that the tool discriminated against
women. Amazon eventually had to abandon the tool in 2017.
• In 2016, ProPublica analyzed a commercially developed system that was created
to help judges make better sentencing decisions by predicting the likelihood that
criminals would reoffend, and it found that it was biased against Black people.10
Incidents such as these are enormous PR nightmares for companies.
Consequently, regulators have published guidelines on the ethical use of data. For
example, EU regulators published a set of seven requirements that must be met for AI
systems to be considered trustworthy:
• AI systems should be under human oversight.
• They need to have a fall-back plan in case something goes wrong. They also need
to be accurate, reliable, and reproducible.
• They must ensure full respect for privacy and data protection.
• Data, system, and AI business models should be transparent and offer
traceability.
• AI systems must avoid unfair bias.
• They must benefit all human beings.
• They must ensure responsibility and accountability.
However, the drive for data-driven decisions, fueled by more data and robust analyt‐
ics, calls for a necessary consideration of and focus on the ethics of data and data use
that goes beyond these regulatory requirements.
Examples of Data Governance in Action
This section takes a closer look at several enterprises and how they were able to derive
benefits from their governance efforts. These examples demonstrate that data gover‐
nance is being used to manage accessibility and security, that it addresses the issue of
trust by tackling data quality head-on, and that the governance structure makes these
endeavors successful.
10 Jonathan Shaw, “Artificial Intelligence and Ethics”, Harvard Magazine, January–February 2019, 44-49, 74.
Examples of Data Governance in Action
|
17
Managing Discoverability, Security, and Accountability
In July 2019, Capital One, one of the largest issuers of consumer and small business
credit cards, discovered that an outsider had been able to take advantage of a miscon‐
figured web application firewall in its Apache web server. The attacker was able to
obtain temporary credentials and access files containing personal information for
Capital One customers.11 The resulting leak of information affected more than 100
million individuals who had applied for Capital One credit cards.
Two aspects of this leak limited the blast radius. First, the leak was of application data
sent to Capital One, and so, while the information included names, social security
numbers, bank account numbers, and addresses, it did not include log-in credentials
that would have allowed the attacker to steal money. Second, the attacker was swiftly
caught by the FBI, and the reason for the attacker being caught is why we include this
anecdote in this book.
Because the files in question were stored in a public cloud storage bucket where every
access to the files was logged, access logs were available to investigators after the fact.
They were able to figure out the IP routes and narrow down the source of the attack
to a few houses. While misconfigured IT systems that create security vulnerabilities
can happen anywhere, attackers who steal admin credentials from on-premises sys‐
tems will usually cover their tracks by modifying the system access logs. On the pub‐
lic cloud, though, these access logs are not modifiable because the attacker doesn’t
have access to them.
This incident highlights a handful of lessons:
• Make sure that your data collection is purposeful. In addition, store as narrow a
slice of the data as possible. It was fortunate that the data store of credit card
applications did not also include the details of the resulting credit card accounts.
• Turn on organizational-level audit logs in your data warehouse. Had this not
been done, it would not have been possible to catch the culprit.
• Conduct periodic security audits of all open ports. If this is not done, no alerts
will be raised about attempts to get past security safeguards.
• Apply an additional layer of security to sensitive data within documents. Social
security numbers, for example, should have been masked or tokenized using an
artificial intelligence service capable of identifying PII data and redacting it.
The fourth best practice is an additional safeguard—arguably, if only absolutely nec‐
essary data is collected and stored, there would be no need for masking. However,
11 “Information on the Capital One Cyber Incident”, Capital One, updated September 23, 2019; Brian Krebs,
“What We Can Learn from the Capital One Hack”, Krebs on Security (blog), August 2, 2019.
18
|
Chapter 1: What Is Data Governance?
most organizations have multiple uses of the data, and in some use cases, the decryp‐
ted social security number might be needed. In order to do such multi-use effectively,
it is necessary to tag or label each attribute based on multiple categories to ensure the
appropriate controls and security are placed on it. This tends to be a collaborative
effort among many organizations within the company. It is worth noting that systems
like these that remove data from consideration come with their own challenges and
risks.12
As the data collected and retained by enterprises has grown, ensuring that best practi‐
ces like these are well understood and implemented correctly has become more and
more important. Such best practices and the policies and tools to implement them are
at the heart of data governance.
Improving Data Quality
Data governance is not just about security breaches. For data to be useful to an orga‐
nization, it is necessary that the data be trustworthy. The quality of data matters, and
much of data governance focuses on ensuring that the integrity of data can be trusted
by downstream applications. This is especially hard when data is not owned by your
organization and when that data is moving around.
A good example of data governance activities improving data quality comes from the
US Coast Guard (USCG). The USCG focuses on maritime search and rescue, ocean
spill cleanup, maritime safety, and law enforcement. Our colleague Dom Zippilli was
part of the team that proved the data governance concepts and techniques behind
what became known as the Authoritative Vessel Identification Service (AVIS). The
following sidebar about AVIS is in his words.
How the US Coast Guard Improved Data Quality
Dom Zippilli
Figure 1-10 illustrates what AVIS looked like when looking at a vessel with no data
discrepancies. The data from Automatic Identification Systems (AIS) corresponds
well with what was in AVIS, which is best described as “what we think we know”
about a ship—an amalgam of information from other USCG systems that handled
vessel registration, integration with the International Maritime Organization (IMO),
citations, and so on.
12 See, for example, the book Dark Data: Why What You Don’t Know Matters by David Hand (Princeton Univer‐
sity Press).
Examples of Data Governance in Action
|
19
Figure 1-10. What AVIS looked like. Figure courtesy of NAVCEN.
Unfortunately, not all data corresponded this cleanly. Figure 1-11 illustrates a patho‐
logical case: no ship image, mismatched name, mismatched Maritime Mobile Service
Identity (MMSI), mismatched IMO number, mismatched everything.
Figure 1-11. A pathological case with a lot of inconsistencies between what’s tracked in
AIS and what’s known from an amalgam of other sources. Figure courtesy of NAVCEN.
Such mismatches make knowing which ships are where, and information about those
ships, much harder to figure out for USCG in the field. The vessels that cropped up in
20
|
Chapter 1: What Is Data Governance?
the AVIS UI were the ones that couldn’t be resolved using automated tooling and thus
required some human intervention. Automating is nice (and this was almost 10 years
ago), but even surfacing the work that had to be done by a human was a huge step
forward. In almost all cases, the issues were the result of innocent mistakes, but get‐
ting things back on track required identifying the issues and reaching out to the mari‐
time community.
The business value of such corrections comes down to maritime domain awareness
(MDA), a key part of the USCG’s mission. Domain awareness is pretty hard to come
by when your data quality is poor. Here are some qualitative examples of how AVIS
helped.
For example, imagine a scenario in which a vessel needs to be investigated for some
kind of violation, or interdicted for any reason. If that vessel is among many broad‐
casting with the same MMSI number, our track for that vessel looks like Figure 1-12.
This could be even more serious in a search and rescue situation in which we need to
locate nearby vessels that could render aid faster than a USCG vessel (cooperation is a
major theme of maritime life).
Figure 1-12. Effects of duplicate vehicle id numbers. Figure courtesy of NAVCEN.
Over time, in the pilot program, as shown in Figure 1-13, we saw a drastic reduction
in the number of ambiguous vessel tracks received each day. While zero was always
the goal, this is by nature a community effort, so it requires constant upkeep.
Examples of Data Governance in Action
|
21
Figure 1-13. Improvements in data quality due to pilot program to correct vessel IDs
The closest I have to a quantitative result (though it doesn’t spell out the mission
value exactly, as that was expected to be obvious to the reader) is this highlight from a
whitepaper that is unfortunately no longer available publicly:
Over the course of the project, the AVIS team was able to virtually eliminate uniden‐
tified and uncorrelated AIS vessel signals broadcasting unregistered MMSI numbers
such as 1, 2, 123456789, etc. Specifically, 863 out of 866 vessels were corrected by
September 2011, eliminating nearly 100% of incorrect broadcasts.13
863 might not seem like a lot, but keep in mind the global merchant fleet is something
on the order of 50,000 vessels. So, just for US waters, this is actually a big part of the
population, and as you know, it doesn’t take a lot of bad data to make all the data
useless.
The USCG program is a handy reminder that data quality is something to strive for
and constantly be on the watch for. The cleaner the data, the more likely it is to be
usable for more critical use cases. In the USCG case, we see this in the usability of the
data for search and rescue tasks as well.
13 David Winkler, “AIS Data Quality and the Authoritative Vessel Identification Service (AVIS)” (PowerPoint
presentation, National GMDSS Implementation Task Force, Arlington, VA, January 10, 2012).
22
|
Chapter 1: What Is Data Governance?
The Business Value of Data Governance
Data governance is not solely a control practice. When implemented cohesively, data
governance addresses the strategic need to get knowledge workers the insights they
require with a clear process to “shop for data.” This makes possible the extraction of
insights from multiple sources that were previously siloed off within different busi‐
ness units.
In organizations where data governance is a strategic process, knowledge workers can
expect to easily find all the data required to fulfill their mission, safely apply for
access, and be granted access to the data under a simple process with clear timelines
and a transparent approval process. Approvers and governors of data can expect to
easily pull up a picture of what data is accessible to whom, and what data is “outside”
the governance zone of control (and what to do about any discrepancies there). CIOs
can expect to be able to review a high-level analysis of the data in the organization in
order to holistically review quantifiable metrics such as “total amount of data” or
“data out of compliance” and even understand (and mitigate) risks to the organiza‐
tion due to data leakage.
Fostering Innovation
A good data governance strategy, when set in motion, combines several factors that
allow a business to extract more value from the data. Whether the goal is to improve
operations, find additional sources of revenue, or even monetize data directly, a data
governance strategy is an enabler of various value drivers in enterprises.
A data governance strategy, if working well, is a combination of process (to make data
available under governance), people (who manage policies and usher in data access
across the organization, breaking silos where needed), and tools that facilitate the
above by applying machine learning techniques to categorize data and indexing the
data available for discovery.
Data governance ideally will allow all employees in the organization to access all data
(subject to a governance process) under a set of governance rules (defined in greater
detail below), while preserving the organization’s risk posture (i.e., no additional
exposure or risks are introduced due to making data accessible under a governance
strategy). Since the risk posture is maintained and possibly even improved with the
additional controls data governance brings, one could argue there is only an upside to
making data accessible. Giving all knowledge workers access to data, in a governed
manner, can foster innovation by allowing individuals to rapidly prototype answers
to questions based on the data that exists within the organization. This can lead to
better decision making, better opportunity discovery, and a more productive organi‐
zation overall.
The Business Value of Data Governance
|
23
The quality of the data available is another way to ascertain whether governance is
well implemented in the organization. A part of data governance is a well-understood
way to codify and inherit a “quality signal” on the data. This signal should tell poten‐
tial data users and analysts whether the data was curated, whether it was normalized
or missing, whether corrupt data was removed, and potentially how trustworthy the
source for the data is. Quality signals are crucial when making decisions on potential
uses of the data; for example, within machine learning training datasets.
The Tension Between Data Governance and Democratizing
Data Analysis
Very often, complete data democratization is thought of as conflicting with data gov‐
ernance. This conflict is not necessarily an axiom. Data democratization, in its most
extreme interpretation, can mean that all analysts or knowledge workers can access
all data, whatever class it may belong to. The access described here makes a modern
organization uncomfortable when you consider specific examples, such as employee
data (e.g., salaries) and customer data (e.g., customer names and addresses). Clearly,
only specific people should be able to access data of the aforementioned types, and
they should do so only within their specific job-related responsibilities.
Data governance is actually an enabler here, solving this tension. The key concept to
keep in mind is that there are two layers to the data: the data itself (e.g., salaries) and
the metadata (data about the data—e.g., “I have a table that contains salaries, but I
won’t tell you anything further”).
With data governance, you can accomplish three things:
• Access a metadata catalog, which includes an index of all the data managed (full
democratization, in a way) and allows you to search for the existence of certain
data. A good data catalog also includes certain access control rules that limit the
bounds of the search (for example, I will be able to search “sales-related data,” but
“HR” is out of my purview completely, and therefore even HR-metadata is inac‐
cessible to me).
• Govern access to the data, which includes an acquisition process (described
above) and a way to adhere to the principle of least access: once access is reques‐
ted, provide access limited to the boundaries of the specific resource; don’t
overshare.
• Independently of the other steps, make an “audit trail” available to the data access
request, the data access approval cycle, and the approver (data steward), as well
as to all the subsequent access operations. This audit trail is data itself and there‐
fore must comply with data governance.
24
|
Chapter 1: What Is Data Governance?
In a way, data governance becomes the facility where you can enable data democra‐
tization, allowing more of your data to be accessible to more of the knowledge
employee population, and therefore be an accelerator to the business in making the
use of data easier and faster.
Business outcomes, such as visibility into all parts of a supply chain, understanding of
customer behavior on every online asset, tracking the success of a multipronged cam‐
paign, and the resulting customer journeys, are becoming more and more possible.
Under governance, different business units will be able to pull data together, analyze
it to achieve deeper insight, and react quickly to both local and global changes.
Manage Risk (Theft, Misuse, Data Corruption)
The key concerns CIOs and responsible data stewards have had for a long time (and
this has not changed with the advent of big data analytics) have always been: What
are my risk factors, what is my mitigation plan, and what is the potential damage?
CIOs have been using these concerns to assign resources based on the answer to
those questions. Data governance comes to provide a set of tools, processes, and posi‐
tions for personnel to manage the risk to data, among other topics presented therein
(for example, data efficiency, or getting value from data). Those risks include:
Theft
Data theft is a concern in those organizations where data is either the product or
a key factor in generating value. Theft of data about parts, suppliers, or price in
an electronics manufacturer supply chain can cause a crippling blow to the busi‐
ness if competition uses that information to negotiate with those very same sup‐
pliers, or to derive a product roadmap from the supply-chain information. Theft
of a customer list can be very damaging to any organization. Setting data gover‐
nance around information that the organization considers to be sensitive can
encourage confidence in the sharing of surrounding data, aggregates, and so on,
contributing to business efficiency and breaking down barriers to sharing and
reusing data.
Misuse
Misuse is often the unknowing use of data in a way that’s different from the pur‐
pose it was collected for—sometimes to support the wrong conclusions. This is
often a result of a lack of information about the data source, its quality, or even
what it means. There is sometimes malicious misuse of data as well, meaning that
information gathered with consent for benign purposes is used for other unin‐
tended and sometimes nefarious purposes. An example is AT&T’s payout to the
FCC in 2015, after its call center employees were found to have disclosed con‐
sumers’ personal information to third parties for financial gain. Data governance
can protect against misuse with several layers. First, establish trust before sharing
data. Another way to protect against misuse is declarative—declare the source of
The Business Value of Data Governance
|
25
the data within the container, the way it was collected, and what it was intended
for. Finally, limiting the length of time for which data is accessible can prevent
possible misuse. This does not mean placing a lid on the data and making it inac‐
cessible. Remember, the fact that the data exists should be shared alongside its
purpose and description—which should make data democratization a reality.
Data corruption
Data corruption is an insidious risk because it is hard to detect and hard to pro‐
tect against. The risk materializes when deriving operational business conclu‐
sions from corrupt (and therefore incorrect) data. Data corruption often occurs
outside of data governance control and can be due to errors on data ingest, join‐
ing “clean” data with corrupt data (creating a new, corrupt product). Partial data,
autocorrected to include some default values, can be misinterpeted, for example,
as curated data. Data governance can step into the fray here and allow recording,
even at the structured data column level, of the processes and lineage of the data,
and the level of confidence, or quality, of the top-level source of the data.
Regulatory Compliance
Data governance is often leveraged when a set of regulations are applicable to the
business, and specifically to the data the business processes. Regulations are, in
essence, policies that must be adhered to in order to play within the business environ‐
ment the organization operates in. GDPR is often referred to as an example regula‐
tion around data. This is because, among other things, GDPR mandates a separation
of (European citizens') personal data from other data, and treatment of that data in a
different way, especially around data that can be used to identify a person. This
manuscript does not intend to go into the specifics of GDPR.
Regulation will usually refer to one or more of the following specifics:
• Fine-grained access control
• Data retention and data deletion
• Audit logging
• Sensitive data classes
Let’s discuss these one by one.
Regulation around fine-grained access control
Access control is already an established topic that relates to security most of all. Finegrained access control adds the following considerations to access control:
26
| Chapter 1: What Is Data Governance?
When providing access, are you providing access to the right size of container?
This means making sure you provide the minimal size of the container of the
data (table, dataset, etc.) that includes the requested information. In structured
storage this will most commonly be a single table, rather than the whole dataset
or project-wide permission.
When providing access, are you providing the right level of access?
Different levels of access to data are possible. A common access pattern is being
able to either read the data or write the data, but there are additional levels: you
can choose to allow a contributor to append (but possibly not change) the data,
or an editor may have access to modify or even delete data. In addition, consider
protected systems in which some data is transformed on access. You could redact
certain columns (e.g., US social security numbers, which serve as a national ID)
to expose just the last four digits, or coarsen GPS coordinates to city and country.
A useful way to share data without exposing too much is to tokenize (encrypt)
the data with symmetric (reversible) encryption such that key data values (for
example, a person’s ID) preserve uniqueness (and thus you can count how many
distinct persons you have in your dataset) without being exposed to the specific
details of a person’s ID.
All the levels of access mentioned here should be considered (read/write/delete/
update and redact/mask/tokenize).
When providing access, for how long should access remain open?
Remember that access is usually requested for a reason (a specific project must be
completed), and permissions granted should not “dangle” without appropriate
justification. The regulator will be asking “who has access to what,” and thus lim‐
iting the number of personnel who have access to a certain class of data will make
sense and can prove efficient.
Data retention and data deletion
A significant body of regulation deals with the deletion and the preservation of data.
A requirement to preserve data for a set period, and no less than that period, is com‐
mon. For example, in the case of financial transaction regulations, it is not uncom‐
mon to find a requirement that all business transaction information be kept for a
duration of as much as seven years to allow financial fraud investigators to backtrack.
Conversely, an organization may want to limit the time it retains certain information,
allowing it to draw quick conclusions while limiting liability. For example, having
constantly up-to-date information about the location of all delivery trucks is useful
for making rapid decisions about “just-in-time” pickups and deliveries, but it
becomes a liability if you maintain that information over a period of time and can, in
theory, plot a picture of the location of a specific delivery driver over the course of
several weeks.
The Business Value of Data Governance
|
27
Audit logging
Being able to bring up audit logs for a regulator is useful as evidence that policies are
complied with. You cannot present data that has been deleted, but you can show an
audit trail of the means by which the data was created, manipulated, shared (and with
whom), accessed (and by whom), and later expired or deleted. The auditor will be
able to verify that policies are being adhered to. Audit logs can serve as a useful foren‐
sic tool as well.
To be useful for data governance purposes, audit logs need to be immutable, writeonly (unchangeable by internal or external parties), and preserved, by themselves, for
a lengthy period—as long as the most demanding data preservation policy (and
beyond that, in order to show the data being deleted).
Audit logs need to include information not only about data and data operations by
themselves but also about operations that happen around the data management
facility. Policy changes need to be logged, and data schema changes need to be logged.
Permission management and permission changes need to be logged, and the logging
information should contain not only the subject of the change (be it a data container
or a person to be granted permission) but also the originator of the action (the
administrator or the service process that initiated the activity).
Sensitive data classes
Very often, a regulator will determine that a class of data should be treated differently
than other data. This is the heart of the regulation that is most commonly concerned
with a group of protected people, or a kind of activity. The regulator will be using
legal language (e.g., personally identifiable data about European Union residents, or
“financial transaction history”). It will be up to the organization to correctly identify
what portion of that data it actually processes, and how this data compares to the data
stored in structured or unstructured storage. For structured data it is sometimes eas‐
ier to bind a data class into a set of columns (PII is stored in these columns) and tag
the columns so that certain policies apply to these columns specifically, including
access and retention. This supports the principles of fine-grained access control as
well as adhering to the regulation about the data (not the data store or the personnel
manipulating that data).
Considerations for Organizations as They Think About Data
Governance
When an organization sits down and begins to define a data governance program and
the goals of such a program, it should take into consideration the environment in
which it operates. Specifically, it should consider what regulations are relevant and
how often these change, whether or not a cloud deployment makes sense for the
28
|
Chapter 1: What Is Data Governance?
organization, and what expertise is required from IT and data analysts/owners. We
discuss these factors next.
Changing regulations and compliance needs
In past years, data governance regulations have garnered more attention. With GDPR
and CCPA joining the ranks of HIPAA- and PCI-related regulations, the affected
organizations are reacting.
The changing regulation environment has meant that organizations need to remain
vigilant when it comes to governance. No organization wants to be in the news for
getting sued for failing to handle customer information as per a set of regulations. In
a world where customer information is very precious, firms need to be careful how
they handle customer data. Not only should firms know about existing regulations,
but they also need to keep up with any changing mandates or stipulations, as well as
any new regulations that might affect how they do business. In addition, changes to
technology have also created additional challenges. Machine learning and AI have
allowed organizations to predict future outcomes and probabilities. These technolo‐
gies also create a ton of new datasets as a part of this process. With these new
predicted values, how do companies think about governance? Should these new data‐
sets assume the same policies and governance that the original datasets had, or
should they have their own set of policies for governance? Who should have access to
this data? How long should it be retained for? These are all questions that need to be
considered and answered.
Data accumulation and organization growth
With infrastructure cost rapidly decreasing, and organizations growing both organi‐
cally and through acquisition of additional business units (with their own data
stores), the topic of data accumulation, and how to properly react to quickly amassing
large amounts of data, becomes important. With data accumulation, an organization
is collecting more data from more sources and for more purposes.
Big data is a term you will keep hearing, and it alludes to the vast amounts of data
(structured and unstructured) now collected from connected devices, sensors, social
networks, clickstreams, and so on. The volume, variety, and velocity of data has
changed and accelerated over the past decade. The effort to manage and even consoli‐
date this data has created data swamps (disorganized and inconsistent collections of
data without clear curation) and even more silos—i.e., customers decided to consoli‐
date on System Applications and Products (SAP), and then they decided to consoli‐
date on Hive Metastore, and some consolidated on the cloud, and so on. Given these
challenges, knowing what you have and applying governance to this data is compli‐
cated, but it’s a task that organizations need to undertake. Organizations thought that
building a data lake would solve all their issues, but now these data lakes are becom‐
ing data swamps with so much data that is impossible to understand and govern. In
The Business Value of Data Governance
|
29
an environment in which IDC predicts that more than a quarter of the data generated
by 2025 will be real-time in nature, how do organizations make sure that they are
ready for this changing paradigm?
Moving data to the cloud
Traditionally, all data resided in infrastructure provided and maintained by the orga‐
nization. This meant the organization had full control over access, and there was no
dynamic sharing of resources. With the emergence of cloud computing—which in
this context implies cheap but shared infrastructure—organizations need to think
about their response and investment in on-premises versus cloud infrastructure.
Many large enterprises still mention that they have no plans to move their core data,
or governed data, to the cloud anytime soon. Even though the largest cloud compa‐
nies have invested money and resources to protect customer data in the cloud, most
customers still feel the need to keep this data on-prem. This is understandable,
because data breaches in the cloud feel more consequential. The potential for damage,
monetary as well as to reputation, explains why enterprises want more transparency
in how governance works to protect their data on the cloud. With this pressure,
you’re seeing cloud companies put more guardrails in place. They need to “show” and
“open the hood” to how governance is being implemented, as well as provide controls
that not only engender trust among customers, but also put some power into custom‐
ers’ hands. We discuss these topics in Chapter 7.
Data infrastructure expertise
Another consideration for organizations is the sheer complexity of the infrastructure
landscape. How do you think about governance in a hybrid and multi-cloud world?
Hybrid computing allows organizations to have both on-premise and cloud infra‐
structure, while multicloud allows organizations to utilize more than one cloud pro‐
vider. How do you implement governance across the organization when the data
resides on-premises and on other clouds? This makes governance complicated and
therefore goes beyond the tools used to implement it. When organizations start
thinking about the people, the processes, and the tools and define a framework that
encompasses these facets, then it becomes a little easier to extend governance across
on-prem and in the cloud.
Why Data Governance Is Easier in the Public Cloud
Data governance involves managing risk. The practitioner is always trading off the
security inherent in never allowing access to the data against the agility that is possi‐
ble if data is readily available within the organization to support different types of
decisions and products. Regulatory compliance often dictates the minimal require‐
ments for access control, lineage, and retention policies. As we discussed in the
30
|
Chapter 1: What Is Data Governance?
previous sections, the implementation of these can be challenging as a result of
changing regulations and organic growth.
The public cloud has several features that make data governance easier to implement,
monitor, and update. In many cases, these features are unavailable or cost-prohibitive
in on-premises systems.
Location
Data locality is mostly relevant for global organizations that store and use data across
the globe, but a deeper look into regulation reveals that the situation is not so simple.
For example, if, for business reasons, you want to leverage a data center in a central
location (say, in the US, next to your potential customers) but your company is a Ger‐
man company, regulation requires that data about employees remains on German
soil; thus your data strategy just became more involved.
The need to store user data within sovereign boundaries is an increasingly common
regulatory requirement. In 2016, the EU Parliament approved data sovereignty meas‐
ures within GDPR, wherein the storage and processing of records about EU citizens
and residents must be carried out in a manner that follows EU law. Specific classes of
data (e.g., health records in Australia, telecommunications metadata in Germany, or
payment data in India) may also be subject to data locality regulations; these go
beyond mere sovereignty measures by requiring that all data processing and storage
occur within the national boundaries. The major public cloud providers offer the
ability to store your data in accordance with these regulations. It can be convenient to
simply mark a dataset as being within the EU multi-region and know that you have
both redundancy (because it’s a multi-region) and compliance (because data never
leaves the EU). Implementing such a solution in your on-premises data center can be
quite difficult, since it can be cost-prohibitive to build data centers in every sovereign
location that you wish to do business in and that has locality regulations.
Another reason that location matters is that secure transaction-aware global access
matters. As your customers travel or locate their own operations, they will require
you to provide access to data and applications wherever they are. This can be difficult
if your regulatory compliance begins and ends with colocating applications and data
in regional silos. You need the ability to seamlessly apply compliance roles based on
users, not just on applications. Running your applications in a public cloud that runs
its own private fiber and offers end-to-end physical network security and global time
synchronization (not all clouds do this) simplifies the architecture of your
applications.
Why Data Governance Is Easier in the Public Cloud
|
31
Reduced Surface Area
In heavily regulated industries, there are huge advantages if there is a single “golden”
source of truth for datasets, especially for data that requires auditability. Having your
enterprise data warehouse (EDW) in a public cloud, particularly in a setting in which
you can separate compute from storage and access the data from ephemeral clusters,
provides you with the ability to create different data marts for different use cases.
These data marts are provided data through views of the EDW that are created on the
fly. There is no need to maintain copies, and examination of the views is enough to
ensure auditability in terms of data correctness.
In turn, the lack of permanent storage in these data marts greatly simplifies their gov‐
ernance. Since there is no storage, complying with rules around data deletion is trivial
at the data mart level. All such rules have to be enforced only at the EDW. Other rules
around proper use and control of the data still have to be enforced, of course. That’s
why we think of this as a reduced surface area, not zero governance.
Ephemeral Compute
In order to have a single source of data and still be able to support enterprise applica‐
tions, current and future, we need to make sure that the data is not stored within a
compute cluster, or scaled in proportion to it. If our business is spiky, or if we require
the ability to support interactive or occasional workloads, we will require infinitely
scalable and readily burstable compute capability that is separate from storage archi‐
tecture. This is possible only if our data processing and analytics architecture is serv‐
erless and/or clearly separates computes and storage.
Why do we need both data processing and analytics to be serverless? Because the util‐
ity of data is often realized only after a series of preparation, cleanup, and intelligence
tools are applied to it. All these tools need to support separation of compute and stor‐
age and autoscaling in order to realize the benefits of a serverless analytics platform.
It is not sufficient just to have a serverless data warehouse or application architecture
that is built around serverless functions. You need your tooling frameworks them‐
selves to be serverless. This is available only in the cloud.
Serverless and Powerful
In many enterprises, lack of data is not the problem—it’s the availability of tools to
process data at scale. Google’s mission of organizing the world’s information has
meant that Google needed to invent data processing methods, including methods to
secure and govern the data being processed. Many of these research tools have been
hardened through production use at Google and are available on Google Cloud as
serverless tools (see Figure 1-14). Equivalents exist on other public clouds as well. For
example, the Aurora database on Amazon Web Services (AWS) and Microsoft’s Azure
32
|
Chapter 1: What Is Data Governance?
Cosmos DB are serverless; S3 on AWS and Azure Cloud Storage are the equivalent of
Google Cloud Storage. Similarly, Lambda on AWS and Azure Functions provide the
ability to carry out stateless serverless data processing. Elastic Map Reduce (EMR) on
AWS and HDInsight on Azure are the equivalent of Google Cloud Dataproc. At the
time of writing, serverless stateful processing (Dataflow on Google Cloud) is not yet
available on other public clouds, but this will no doubt be remedied over time. These
sorts of capabilities are cost-prohibitive to implement on-premises because of the
necessity to implement serverless tools in an efficient manner while evening out the
load and traffic spikes across thousands of workloads.
Figure 1-14. Many of the data-processing techniques invented at Google (top panel; see
also http://research.google.com/pubs/papers.html) exist as managed services on Google
Cloud (bottom panel).
Labeled Resources
Public cloud providers provide granular resource labeling and tagging in order to
support a variety of billing considerations. For example, the organization that owns
the data in a data mart may not be the one carrying out (and therefore paying for) the
compute. This gives you the ability to implement regulatory compliance on top of the
sophisticated labeling and tagging features of these platforms.
Why Data Governance Is Easier in the Public Cloud
|
33
These capabilities might include the ability to discover, label, and catalog items (ask
your cloud provider whether this is the case). It is important to be able to label
resources, not just in terms of identity and access management but also in terms of
attributes, such as whether a specific column is considered PII in certain jurisdic‐
tions. Then it is possible to apply consistent policies to all such fields everywhere in
your enterprise.
Security in a Hybrid World
The last point about having consistent policies that are easily applicable is key. Con‐
sistency and a single security pane are key benefits to hosting your enterprise soft‐
ware infrastructure on the cloud. However, such an all-or-nothing approach is
unrealistic for most enterprises. If your business operates equipment (handheld devi‐
ces, video cameras, point-of-sale registers, etc.) “on the edge,” it is often necessary to
have some of your software infrastructure there as well. Sometimes, as with voting
machines, regulatory compliance might require physical control of the equipment
being used. Your legacy systems may not be ready to take advantage of the separation
of compute and storage that the cloud offers. In these cases, you’d like to continue to
operate on-premises. Systems that involve components that live in a public cloud and
one other place—in two public clouds, or in a public cloud and on the edge, or a in
public cloud and on-premises—are termed hybrid cloud systems.
It is possible to greatly expand the purview of your cloud security posture and poli‐
cies by employing solutions that allow you to control both on-premises and cloud
infrastructure using the same tooling. For example, if you have audited an onpremises application and its use of data, it is easier to approve that identical applica‐
tion running in the cloud than it is to reaudit a rewritten application. The cost of
entry to this capability is to containerize your applications, and this might be a cost
well worth paying for, for the governance benefits alone.
Summary
When discussing a successful data-governance strategy, you must consider more than
just the data architecture/data pipeline structure or the tools that perform “gover‐
nance” tasks. Consideration of the actual humans behind the governance tools as well
as the “people processes” put into place is also highly important and should not be
discounted. A truly successful governance strategy must address not only the tools
involved but the people and processes as well. In Chapters 2 and 3, we will discuss
these ingredients of data governance.
In Chapter 4, we take an example corpus of data and consider how data governance is
carried out over the entire life cycle of that data; from ingest to preparation and stor‐
age, to incorporation into reports, dashboards, and machine learning models, and on
to updates and eventual deletion. A key concern here is that data quality is an
34
|
Chapter 1: What Is Data Governance?
ongoing concern; new data-processing methods are invented, and business rules
change. How to handle the ongoing improvement of data quality is addressed in
Chapter 5.
By 2025, more than 25% of enterprise data is expected to be streaming data. In Chap‐
ter 6, we address the challenges of governing data that is on the move. Data in flight
involves governing data at the source and at the destination, and any aggregations
and manipulations that are carried in flight. Data governance also has to address the
challenges of late-arriving data and what it means for the correctness of calculations if
storage systems are only eventually correct.
In Chapter 7, we delve into data protection and the solutions available for authentica‐
tion, security, backup, and so on. The best data governance is of no use if monitoring
is not carried out and leaks, misuse, and accidents are not discovered early enough to
be mitigated. Monitoring is covered in Chapter 8.
Finally, in Chapter 9, we bring together the topics in this book and cover best practi‐
ces in building a data culture—a culture in which both the user and the opportunity
is respected.
One question we often get asked is how Google does data governance internally. In
Appendix A, we use Google as an example (one that we know well) of a data gover‐
nance system, and point out the benefits and challenges of the approaches that Goo‐
gle takes and the ingredients that make it all possible.
Summary
|
35
CHAPTER 2
Ingredients of Data Governance: Tools
A lot of the tasks related to data governance can benefit from automation. Machine
learning tools and automatic policy applications or suggestions can accelerate data
governance tasks. In this chapter we will review some of the tools commonly referred
to when discussing data governance.
When evaluating a data governance process/system, pay attention to the capabilities
mentioned in this chapter. The following discussion concerns tasks and tools that can
augment complete end-to-end support for the processes involved in, and the person‐
nel responsible for, data governance organization. We’ll dive into the various pro‐
cesses and solutions in more detail in later chapters.
The Enterprise Dictionary
To begin, it is important to understand how an organization works with data and
enables data governance. Usually, there is an enterprise dictionary or an enterprise
policy book of some kind.
The first of these documents, the enterprise dictionary, is one that can take many
shapes, from a paper document to a tool that encodes or automates certain policies. It
is an agreed-upon repository of the information types (infotypes) used by the organi‐
zation—that is, data elements that the organization processes and derives insights
from. An infotype will be a piece of information with a singular meaning—“email
address” or “street address,” for example, or even “salary amount.”
In order to refer to individual fields of information and drive a governance policy
accordingly, you need to name those pieces of information.
37
An organization’s enterprise dictionary is normally owned by either the legal depart‐
ment (whose focus would be compliance) or the data office (whose focus would be
standardization of the data elements used).
In Figure 2-1, you can see examples of infotypes, and a potential organization of
those infotypes into organization-specific data classes to be used in policy making.
Figure 2-1. Data classes and infotypes
Once the enterprise dictionary is defined, the various individual infotypes within it
can be grouped into data classes, and a policy can be defined for each data class. The
enterprise dictionary generally contains data classes, policies related to the data
classes, and additional metadata. The following sections expand on this.
Data Classes
A good enterprise dictionary will contain a listing of the classes of data the organiza‐
tion processes. Those will be infotypes collected into groups that are treated in a com‐
mon way from the policy management aspect. For example, an organization will not
want to treat “street addresses,” “phone numbers,” “city, state,” and “zip code” differ‐
ently in a granular manner but must rather be able to set a policy that “all location
information for consumers must be accessible only to a privileged group of personnel
and be kept only for a maximum of 30 days.” This means that the enterprise dictio‐
nary we’ve just described, will actually contain a hierarchy of infotypes—at the leaf
nodes there will be the individual infotypes (e.g., “address,” “email”), and at the root
nodes you will find a data class, or a sensitivity classification (or sometimes both).
Figure 2-2 shows an example of such a hierarchy from a fictional organization.
38
|
Chapter 2: Ingredients of Data Governance: Tools
Figure 2-2. A data class hierarchy
In the data class hierarchy detailed in Figure 2-2, you can see how infotypes such as
IMEI (cellular device hardware ID), phone number, and IP address are grouped
together under personally identifiable information (PII). For this organization, these
are easily identifiable automatically, and policies are defined on “all PII data ele‐
ments.” PII is paired with PHI (protected health information) in the “restricted data”
category. It is likely that there are further policies defined on all data grouped under
the “restricted” heading.
Data classes are usually maintained by a central body within the organization,
because policies on “types of data classes” usually affect compliance to regulation.
Some example data classes seen across many organizations are:
The Enterprise Dictionary
|
39
PII
This is data such as name, address, and personal phone number that can be used
to uniquely identify a person. For a retailer, this can be a customer list. Other
examples include lists of employee data, a list of third-party vendors, and similar
information.
Financial information
This is data such as transactions, salaries, benefits, or any kind of data that can
include information of financial value.
Business intellectual property
This is information related to the success and differentiation of the business.
The variety and kind of data classes will change with the business vertical and inter‐
est. The preceding example data classes (PII, financial, business intellectual property)
will not work for every organization, and the meaning of data collected and the clas‐
sification will differ between businesses and between countries. Do note that data
classes are a combination of information elements belonging to one topic. For exam‐
ple, a phone number is usually not a data class, but PII (of which phone number is a
member) is normally a data class.
The characteristics of a data class are twofold:
• A data class references a set of policies: the same retention rules and access rules
are required on this data.
• A data class references a set of individual infotypes, as in the example in
Figure 2-1.
Data Classes and Policies
Once the data the organization handles is defined in an enterprise dictionary, policies
that govern the data classes can be assigned to data across multiple containers. The
desired relationship is between a policy (e.g., access control, retention) and a data class
rather than a policy and an individual container. Associating policies with the mean‐
ing of data allows for better human understanding (“Analysts cannot access PII” ver‐
sus “Analysts cannot access column #3 on Table B”) and supports scaling to larger
amounts of data.
Above, we have discussed the relationship between data classes and policies. Fre‐
quently, along with the data class specification, the central data office, or legal, will
define an enterprise policy book. An organization needs to be able to answer the
question, “What kinds of data do we process?” An enterprise policy book records
that. It specifies the data classes the organization uses, the kinds of data that are
40
| Chapter 2: Ingredients of Data Governance: Tools
processed, and how they are processed, and it elaborates on “what are we allowed and
not allowed to do” with the data. This is a crucial element in the following respects:
• For compliance, the organization needs to be able to prove to a regulator that it
has the right policies in place around the handling of data.
• A regulator will require the organization to submit the policy book, as well as
proof (usually from audit logs) of compliance with the policies.
• The regulator will require evidence of procedures to ensure that the policy book
is enforced and may even comment on the policies themselves.
We’d like to stress the importance of having a well-thought-out and
well-documented enterprise policy book. Not only is it incredibly
helpful for understanding, organizing, and enforcing your policies,
but it also enables you to quickly and effectively react to changing
requirements and regulations. Additionally, it should be noted that
having the ability to quickly and easily provide documentation and
proof of compliance—not only for external audits but also for
internal ones—should not be discounted. Knowing at a glance how
your company is doing with its policies and its management of
those policies is critical for ensuring a successful and comprehen‐
sive governance program. During the course of our research and
interviews, many companies expressed their struggles with being
able to conduct quick internal audits to know how their gover‐
nance strategy was going. This often resulted in much time and
effort being spent backtracking and documenting policies, how
they were enforced, on what data they were attached, and so on.
Creating an enterprise policy book and setting yourself up to more
easily audit internally (and thus externally) will help you avoid this
pain.
To limit liability, risk management, and exposure to legal action, an organization will
usually define a maximum (and a minimum) retention rate for data. This is impor‐
tant because during an investigation, certain law enforcement agencies will require
certain kinds of data, which the organization must therefore be able to supply. In the
case of financial institutions, for example, it is common to find requirements for
holding certain kinds of data (e.g., transactions) for a minimum of seven years. Other
kinds of data pose a liability: you cannot leak or lose control of data that you don’t
have.
Another kind of policy will be access control. For data, access control goes beyond
“yes/no” and into no access, partial access, or full access. Partial access is accessing the
data when some bits have been “starred out,” or accessing the data after a determinis‐
tic encryption transformation, which will still allow acting on distinct values, or
The Enterprise Dictionary
|
41
grouping by these values, without being exposed to the underlying cleartext. You can
think of partial access as a spectrum of access, ranging from zero access to
ever-increasing details about the data in question (format only, number of digits only,
tokenized rendition, to full access). See Figure 2-3.
Figure 2-3. Examples of varying levels of access for sensitive data
Normally, a policy book will specify:
• Who (inside or outside the organization) can access a data class
• The retention policy for the data class (how long data is preserved)
• Data residency/locality rules, if applicable
• How the data can be processed (OK/Not OK for analytics, machine learning,
etc.)
• Other considerations by the organization
The enterprise policy book—and, with it, the enterprise dictionary—describe the data
managed by the organization. Now let’s discuss specific tools and functionality that
can accelerate data governance work and optimize personnel time.
Per-Use-Case Data Policies
Data can have different meanings and require different policies when the data use
case is taken into consideration. An illustrative example might be a furniture manu‐
facturer that collects personal data (names, addresses, contact numbers) in order to
ensure delivery. The very same data can potentially be used for marketing purposes,
but consent, in most cases, probably has not been granted for the purpose of market‐
ing. However, at the same time, I would very much like that sofa to be delivered to my
home, so I want the manufacturer to store the data! The use case, or purpose, of the
data access ideally should be an overlay on top of your organizational membership
and organizational roles. One way to think about this would be as a “window”
through which the analyst can select data, specifying the purpose ahead of time, and
potentially moving the data into a different container for that purpose (the marketing
42
|
Chapter 2: Ingredients of Data Governance: Tools
database, for example), all with an audit artifact and lineage tracking that will be used
for tracking purposes.
Importance of Use Case and Policy Management
Increasingly, as requirements and regulations change to accommodate the new and
different types of data that are collected, the use case of data becomes an important
aspect of policy management. One of the things that we’ve heard companies say over
and over is that the use case of data needs to be a part of policies—the simple “data
class to role-based access type” isn’t sufficient. We’ve given the example of a furniture
manufacturer and access to, and usage of, customer address for delivery versus mar‐
keting purposes. Another example we’ve heard about throughout our research is that
of a company that has employees who may perform multiple roles. In this case, a sim‐
ple role-based access policy is inadequate, because these employees may be allowed
access to a particular set of data for performing tasks relating to one role but may not
have access for tasks relating to a different role. From this you can see how it’s more
efficient (and effective) to consider access related to what the data will be used for—its
use case—rather than only considering infotype/data class and employee role.
Data Classification and Organization
To control the governance of data, it is beneficial to automate, at least in part, the
classification of data into, at the very least, infotypes—although an even greater auto‐
mation is sometimes adopted. A data classifier will look at unstructured data, or even
a set of columns in structured data, and infer “what” the data is—e.g., it will identify
various representations of phone numbers, bank accounts, addresses, location indica‐
tors, and more.
An example classifier would be Google’s Cloud Data Loss Prevention (DLP). Another
classifier is Amazon’s Macie service.
Automation of data classification can be accomplished in two main ways:
• Identify data classes on ingest, triggering a classification job on the addition of
data sources
• Trigger a data-classification job periodically, reviewing samples of your data
When it is possible, identifying new data sources and classifying them as they are
added to the data warehouse is most efficient, but sometimes this is not possible with
legacy or federated data.
Upon classifying data, you can do the following, depending on the desired level of
automation:
The Enterprise Dictionary
|
43
• Tag the data as “belonging to a class” (see “The Enterprise Dictionary” on page
37)
• Automatically (or manually) apply policies that control access to and retention of
the data according to the definition of the data class, “purpose,” or context for
which the data is accessed or manipulated
Data Cataloging and Metadata Management
When talking about data, data classification, and data classes, we need to discuss the
metadata, or the “information about information”—specifically, where it’s stored and
what governance controls there are on it. It would be naive to think that metadata
obeys the same policies and controls as the underlying data itself. There are many
cases, in fact, where this can be a hindrance. Consider, for example, searching in a
metadata catalog for a specific table containing customer names. While you may not
have access to the table itself, knowing such a table exists is valuable (you can then
request access, you can attempt to review the schema and figure out if this table is
relevant, and you can avoid creating another iteration of this information if it already
exists). Another example is data residency–sensitive information, which must not
leave a certain national border, but at the same time that restriction does not neces‐
sarily apply to information about the existence of the data itself, which may be rele‐
vant in a global search. A final example is information about a listing of phone calls
(who called whom, from where, and when), which can be potentially more sensitive
than the actual calls themselves, because a call list places certain people at certain
locations at certain times.
Crucial to metadata management is a data catalog, a tool to manage this metadata.
Whereas enterprise data warehouses, such as Google BigQuery, are efficient at pro‐
cessing data, you probably want a tool that spans multiple storage systems to hold the
information about the data. This includes where the data is and what technical infor‐
mation is associated with it (table schema, table name, column name, column
description)—but you also should allow for the attachment of additional “business”
metadata, such as who in the organization owns the data, whether the data is locally
generated or externally purchased, whether it relates to production use cases or test‐
ing, and so on.
As your data governance strategy grows, you will want to attach the particulars of
data governance information to the data in a data catalog: data class, data quality,
sensitivity, and so on. It is useful to have these dimensions of information schema‐
tized, so that you can run a faceted search like “Show me all data of type:table and
class:X in the “production” environment”.
44
|
Chapter 2: Ingredients of Data Governance: Tools
The data catalog clearly needs to efficiently index all this information and must be
able to present it to the users whose permissions allow it, using high-performing
search and discovery tooling.
Data Assessment and Profiling
A key step in most insight generation workflows, as you sift through data, is to review
that data for outliers. There are many possible reasons for outliers, and best practice
is to review before making sweeping/automated decisions. Outliers can be the result
of data-entry errors or may just be inconsistent with the rest of the data, but they can
also be weak signals or less-represented new segments or patterns. In many cases, you
will need to normalize the data for the general case before driving insights. This nor‐
malization (keeping or removing outliers) should be done in the context of the busi‐
ness purpose the data is being used for—for example, if you are looking for unusual
patterns or not.
The reason for normalizing data is to ensure data quality and consistency (sometimes
data entry errors lead to data inconsistencies). Again, this must be done in the con‐
text of the business purpose of the data use case. Data quality cannot be performed
without a directing business case, because “reviewing transactions” is not the same
for a marketing team (which is looking for big/influential customers) and a fraud
analysis team (which is looking for provable indications of fraud).
Note that machine learning models, for example, are susceptible to arriving at the
wrong conclusions by extracting generalizations from erroneous data. This is also
true for many other types of analytics.
Data engineers are usually responsible for producing a report that contains data outli‐
ers and other suspected quality issues. Unless the data errors are obviously caused by
data entry or data processing, the data errors are fixed/cleansed by data analysts who
are part of the organization that owns/produces the data, as previously mentioned.
This is done in light of the expected use of the data. However, the data engineers can
leverage a programmatic approach in fixing the data errors, as per the data owners’
requests and requirements. The data engineers will look for empty fields, out-ofbounds values (e.g., people of ages over 200 or under 0), or just plain errors (a string
where a number is expected). There are tools to easily review a sample of the data and
make the cleanup process easier, for example, Dataprep by Trifacta and Stitch.
These cleanup processes work to ensure that using the data in applications, such as
generating a machine learning model, does not result in it being skewed by data outli‐
ers. Ideally, data should be profiled in order to detect anomalies per column and
make a determination on whether anomalies are making sense in the relevant context
(e.g., customers shopping in a physical store outside of store hours are probably an
error, while late-night online ordering is very much a reality). The bounds for what
The Enterprise Dictionary
|
45
kinds of data are acceptable for each field are set, and automated rules prepare and
clean up any batch of data, or any event stream, for ingestion. Care must be taken to
avoid introducing bias into the data, such as by eliminating outliers where they
should not be eliminated.
Data Quality
Data quality is an important parameter in determining the relevant use cases for a
data source, as is being able to rely on data for further calculations/inclusions with
other datasets. You can identify data quality by looking at the data source—i.e.,
understanding where it physically came from. (Error-prone human entry? Fuzzy IoT
devices optimizing for quantity, not quality? Highly exact mobile app event stream?)
Knowing the quality of data sources should guide the decision of whether to join
datasets of varying quality, because low-quality data will reduce confidence in higherquality sources. Data quality management processes include creating controls for val‐
idation, enabling quality monitoring and reporting, supporting the triage process for
assessing the level of incident severity, enabling root cause analysis and recommenda‐
tion of remedies to data issues, and data incident tracking.
There should be different confidence levels assigned to different quality datasets.
There should also be considerations around allowing (or at least curating) resultant
datasets with mixed-quality ancestors. The right processes for data quality manage‐
ment will provide measurably trustworthy data for analysis.
One possible process that can be implemented to improve data quality is a sense of
ownership: making sure the business unit responsible for generating the data also
owns the quality of that data and does not leave it behind for users downstream. The
organization can create a data acceptance process wherein data is not allowed to be
used until the owners of the data prove the data is of a quality that passes the organi‐
zation’s quality standards.
Lineage Tracking
Data does not live in a vacuum; it is generated by certain sources, undergoes various
transformations, aggregates additionals, and is eventually supporting certain insights.
A lot of valuable context is generated from the source of the data and how it was
manipulated along the way, which is crucial to track. This is data lineage.
Why is lineage tracking important? One reason is understanding the quality of a
resulting dashboard/aggregate. If that end product was generated from high-quality
data, but later the information is merged into lower-quality data, that leads to a differ‐
ent interpretation of the dashboard. Another example will be viewing, in a holistic
46
|
Chapter 2: Ingredients of Data Governance: Tools
manner, the movement of a sensitive data class across the organization data scape, to
make sure sensitive data is not inadvertently exposed into unauthorized containers.
Lineage tracking should be able, first and foremost, to present a calculation on the
resultant metrics, such as “quality,” or on whether or not the data was “tainted” with
sensitive information. And later it must be able to present a graphical “graph” of the
data traversal itself. This graph is very useful for debugging purposes but is less so for
other purposes.
Lineage Tracking and Time/Cost
When describing how lineage tracking—especially doing so visually—is so important,
companies have often referred to the level of effort they have to put into debugging
and troubleshooting. They have stated how much time is spent not on fixing prob‐
lems or errors but just on trying to find out if there are any errors or problems at all.
We have heard time and time again that better tracking (e.g., notifications about
what’s going on and when things are “on fire” and should be looked at) not only
would help solve issues better but also would save valuable time—and thus expense—
in the long run. Often when lineage is talked about, the focus is on knowing where
data came from and where it’s going to, but there is also value in visually seeing/
knowing when and where something breaks and being able to take immediate action
on it.
Lineage tracking is also important when thinking about explaining decisions later on.
By identifying input information into a decision-making algorithm (think about a
neural net, or a machine learning model), you can rationalize later why some busi‐
ness decisions (e.g., loan approval) were made in a certain way in the past and will be
made in a certain way in the future. By making business decisions explainable (past
transactions explaining a current decision) and keeping this information transparent
to the data users, you practice good data governance.
This also brings up the importance of the temporal dimension of lineage. The more
sophisticated solutions track lineage across time—tracking not only what the current
inputs to a dashboard are but also what those inputs were in the past, and how the
landscape has evolved.
Key Management and Encryption
One consideration when storing data in any kind of system is whether to store it in a
plain-text format or whether to encrypt it. Data encryption provides another layer of
protection (beyond protecting all data traffic itself), as only the systems or users that
The Enterprise Dictionary
|
47
have the keys can derive meaning from the data. There are several implementations
of data encryptions:
• Data encryption where the underlying storage can access the key. This allows the
underlying storage system to effect efficient storage via data compression
(encrypted data usually does not compress well). When the data is accessed out‐
side the bounds of the storage system—for example, if a physical disk is taken out
of a data center—the data should be unreadable and therefore secure.
• Data encryption where the data is encrypted by a key inaccessible to the storage
system, usually managed separately by the customer. This provides protection
from a bad actor within the storage provider itself in some cases but results in
inefficient storage and performance impact.
• Just-in-time decryption, where, in some cases and for some users, it is useful to
decrypt certain data as it is being accessed as a form of access control. In this
case, encryption works to protect some data classes (e.g., “customer name”) while
still allowing insights such as “total aggregate revenues from all customers” or
“top 10 customers by revenue,” or even identifying subjects who meet some con‐
dition, with the option to ask for de-masking of these subjects later via a trouble
ticket.
All data in Google Cloud is encrypted by default, both in transit and at rest, ensuring
that customer data is always protected from intrusions and attacks. Customers can
also choose customer-managed encryption keys (CMEK) using Cloud KMS, or they
can opt for customer-supplied encryption keys (CSEK) when they need more control
over their data.
To provide the strongest protections, your encryption options should be native to the
cloud platform or data warehouse you choose. The big cloud platforms all have a
native key management that usually allows you to perform operations on keys
without revealing the actual keys. In this case, there are actually two keys in play:
A data encryption key (DEK)
This key is used to directly encrypt the data by the storage system.
A key encryption key (KEK)
This key is used to protect the data encryption key and resides within a protected
service, a key management service.
48
| Chapter 2: Ingredients of Data Governance: Tools
A Sample Key Management Scenario
In the scenario depicted in Figure 2-4, the table on the right is encrypted in chunks
using the plain data encryption key.1 The data encryption key is not stored with the
table but instead is stored in a protected form (wrapped) by a striped key encryption
key. The key encryption key resides (only) in the key management service.
Figure 2-4. Key management scenario
To access the data, a user (or process) follows these steps:
1. The user/process requests the data, instructing the data warehouse (BigQuery) to
use the “striped key” to unwrap the data encryption key, basically passing the key
ID.
2. BigQuery retrieves the protected DEK from the table metadata and accesses the
key management service, supplying the wrapped key.
3. The key management service unwraps the data protection key, while the KEK
never leaves the vault of the key management service.
4. BigQuery uses the DEK to access the data and then discards it, never storing it in
a persistent manner.
This scenario ensures that the key encryption key never leaves a secure separate store
(the KMS) and that the data encryption key never resides on disk—only in memory,
and only when needed.
1 Protection of data at rest is a broad topic; a good starter book would be Applied Cryptography by Bruce Schne‐
ier (John Wiley and Sons).
The Enterprise Dictionary
|
49
Data Retention and Data Deletion
Another important item in the data governance tool chest is the ability to control
how long data is kept—that is, setting maximal and minimal values. There are clear
advantages to identifying data that should survive occasional storage space optimiza‐
tion as being more valuable to retain, but setting a maximum amount of time on data
retention for a less-valuable data class and automatically deleting it seems less obvi‐
ous. Consider that retaining PII presents the challenges of proposer disclosure,
informed consent, and transparency. Getting rid of PII after a short duration (e.g.,
retaining location only while on the commute) simplifies the above.
When talking about data retention and data deletion, we’re often thinking about them
in the context of how to treat sensitive data—that is, whether to retain it, encrypt it,
delete it, or handle it some other way. But there are also scenarios around which your
governance policies not only can save you from being out of compliance but also can
protect you from lost work.
Although it’s a bit old, an example that comes to mind around the subject of protec‐
tion against data deletion is the accidental deletion of the movie Toy Story 2 in 1998.2
During the film’s production process, Oren Jacob, Pixar’s CTO, and Technical Direc‐
tor Galyn Susman were looking at a directory that was holding the assets for several
characters when they encountered an error—something along the lines of “directory
no longer valid,” meaning the location of those characters in the directory had been
deleted. During their effort to walk back through the directory to find where the
problem occurred, they witnessed, in real time, assets for several of the characters
vanish before their eyes.
When all was said and done, an erroneous 10-character command had mistakenly
deleted 90% of the movie. They were able to recover most of the movie, but unfortu‐
nately about a week’s worth of work was lost forever.
You can see from this example that even though sensitive data wasn’t leaked or lost, a
multitude of data was still accidentally deleted—some of which could never be recov‐
ered. We challenge you to consider in your governance program not only how you
will deal with and treat sensitive data in terms of where or how long you retain it, and
whether or not you delete it, but also how that same program can be implemented on
other classes and categories of data that are important for you to back up. While a loss
of data may not result in a breach in compliance, it could certainly result in other
catastrophic business consequences that, if planned for, will hopefully not happen to
you.
2 Dan Moshkovich, “Accidental Data Deletion: The Cautionary Tale of Toy Story 2 and MySpace”, HubStor
(blog), August 17, 2020.
50
| Chapter 2: Ingredients of Data Governance: Tools
Case Study: The Importance of Clear Data Retention Rules for
Data Governance
A lesson about data retention was learned by a Danish taxi company that actually had
a data governance policy in place.3
To understand the story, we need a little bit of context about GDPR Article 5, the reg‐
ulation covering treatment of European citizens’ data by tech companies. GDPR
details the standards that organizations have to follow when processing personal data.
These standards state that data must be handled in a way that is transparent to its sub‐
ject (the European citizen), collected for a specific purpose, and used only for that
purpose. GDPR Article 5(1)(c) addresses data minimization by requiring that per‐
sonal data be limited to what is necessary relative to the purpose for which it is pro‐
cessed. And Article 5(1)(e)—the one that’s most relevant to the taxi company example
—specifies that data cannot be stored any longer than is necessary for the purposes
for which it was gathered.
The Danish taxi company in this example (Figure 2-5) processed taxi ridership data
for legitimate purposes, making sure there was a record of every ride, and of the asso‐
ciated fare, for example, for chargeback and accounting purposes.
Figure 2-5. Danish taxis (photo courtesy of Håkan Dahlström, Creative Commons
License 2.0
The Danish taxi company, as mentioned, had a data retention policy in place: after
two years, the data from a taxi ride was made anonymous by deleting the name of the
passenger. However, that action did not make the data completely anonymous, as mul‐
tiple additional details were (albeit transparently) collected. These included the geolo‐
cation of the taxi ride’s start and end and the phone number of the passenger. With
these details, even without the name of the passenger, it was easy to identify the
3 For more about this case study, see coverage by the law firm Gorrissen Federspiel.
The Enterprise Dictionary
|
51
passenger, and therefore the company was not in compliance with its own statement
of anonymization; thus the data retention policy was actually not effective.
The lesson learned is that considering the business goal of a data retention policy, and
whether or not that goal is actually achieved by the policy set in place, is essential. In
our case, the taxi company was fined by the European Union and was criticized for
the fact that telephone numbers are actually used as unique identifiers within the taxi
ride management system.
Workflow Management for Data Acquisition
One of the key workflows tying together all the tools mentioned so far is data acquisi‐
tion. This workflow usually begins with an analyst seeking data to perform a task.
The analyst, through the power of a well-implemented data governance plan, is able
to access the data catalog for the organization and, through a multifaceted search
query, is able to review relevant data sources. Data acquisition continues with identi‐
fying the relevant data source and seeking an access grant to it. The governance con‐
trols send the access request to the right authorizing personnel, and access is granted
to the relevant data warehouse, enforced through the native controls of that ware‐
house. This workflow—identifying a task, shopping for relevant data, identifying rele‐
vant data, and acquiring access to it—constitutes a data access workflow that is safe.
The level of data access requested—that is, access to metadata for search, access to the
data itself, querying the data in aggregate—these are all data governance decisions.
IAM—Identity and Access Management
When talking about data acquisition, it’s important to detail how access control
works. The topic of access control relies on user authentication and, per the user, the
authorization of the user to access certain data and the conditions of access.
The objective of authenticating a user is to determine that “you are who you say you
are.” Any user (and, for that matter, any service or application) operates under a set of
permissions and roles tied to the identity of a service. The importance of securely
authenticating a user is clear: if I can impersonate a different user, there is a risk of
assuming that user’s roles and privileges and breaking data governance.
Authentication used to be traditionally accomplished by supplying a password tied to
the user requesting access. This method has the obvious drawback that anyone who
has somehow gained access to the password can gain access to everything that user
has access to. Nowadays, proper authentication requires:
52
|
Chapter 2: Ingredients of Data Governance: Tools
• Something you know—this will be your password or passphrase; it should be
hard to guess and changed regularly.
• Something you have—this serves as a second factor of authentication. After pro‐
viding the right passphrase, a user will be prompted to prove that they have a cer‐
tain device (a cell phone able to accept single-use codes, a hardware token),
adding another layer of security. The underlying assumption is that if you mis‐
place that “object” you would report it promptly, ensuring the token cannot be
used by others.
• Something you are—sometimes, for another layer of security, the user will add
biometric information to the authentication request: a fingerprint, a facial scan,
or something similar.
• Additional context—another oft-used layer of security is ensuring that an
authenticated user can access only certain information from within a specific
sanctioned application or device, or other conditions. Such additional context
often includes:
— Being able to access corporate information only from corporate hardware
(sanctioned and cleared by central IT). This, for example, eliminates the risk
of “using the spouse’s device to check for email” without enjoying the default
corporate anti-malware software installed by default on corporate hardware.
— Being able to access certain information only during working hours—thus
eliminating the risk of personnel using their off-hours time to manipulate
sensitive data, maybe when those employees are not in appropriate surround‐
ings or are not alert to risk.
— Having limited access to sensitive information when not logged in to the cor‐
porate network—when using an internet café, for example, and risking net‐
work eavesdropping.
The topic of authentication is the cornerstone of access control, and each organiza‐
tion will define its own balance between risk aversion and user-authentication fric‐
tion. It is a known maxim that the more “hoops” employees have to jump through in
order to access data, the more these employees will seek to avoid complexity, leading
to shadow IT and information siloing—directions opposed to data governance (data
governance seeks to promote data access to all, under proper restrictions). There are
detailed volumes written on this topic.4
4 An example book about identity and access management is Identity and Access Management by Ertem Osma‐
noglu (Elsevier Science, Syngress).
The Enterprise Dictionary
|
53
User Authorization and Access Management
Once the user is properly authenticated, access is determined by a process of checking
whether the user is authorized to access or otherwise perform an operation on the
data object in question, be it a table, a dataset, a pipeline, or streaming data.
Data is a rich medium, and sample access policies can be:
• For reading the data directly (performing “select” SQL statement on a table, read‐
ing a file).
• For reading or editing the metadata associated with the data. For a table, this
would be the schema (the names and types of columns, the table name). For a
file, this would be the filename. In addition, metadata also refers to the creation
date, update date, and “last read” dates.
• For updating the content, without adding new content.
• For copying the data or exporting it.
• There are also access controls associated with workflows, such as performing an
extract-transform-load (ETL) operation to move and reshape the data (replacing
rows/columns with others).
We have expanded here on the policies previously mentioned for data classes, which
also detail partial-read access—which can be its own authorized function.
It’s important to define identities, groups, and roles and assign access rights to estab‐
lish a level of managed access.
Identity and access management (IAM) should provide role management for every
user, with the capability to flexibly add custom roles that group together meaningful
permissions relevant to your organization, ensuring that only authorized and authen‐
ticated individuals and systems are able to access data assets according to defined
rules. Enterprise-scale IAM should also provide context (IP, device, the time the
access request is being generated from, and, if possible, use case for access). As good
governance results in context-specific role and permission determination before any
data access, the IAM system should scale to millions of users, issuing multiple data
access requests per second.
54
|
Chapter 2: Ingredients of Data Governance: Tools
Summary
In this chapter, we have gone through the basic ingredients of data governance: the
importance of having a policy book containing the data classes managed, and how to
clean up the data, secure it, and control access. Now it is time to go beyond the
tooling and discuss the importance of the additional ingredients of people and pro‐
cesses to a successful data governance program.
Summary
|
55
CHAPTER 3
Ingredients of Data Governance:
People and Processes
As mentioned in the preceding chapters, companies want to be able to derive more
insights from their data. They want to make “data-driven decisions.” Gone are the
days of business decisions being based exclusively on intuition or observation. Big
data and big data analytics now allow for decisions to be made based on collecting
data and extracting patterns and facts from that data.
We have spent much time explaining how this movement into using big data has
brought with it a host of considerations around the governance of that data, and we
have outlined tools that aid in this process. Tools, however, are not the only factor to
evaluate when designing a data governance strategy—the people involved and the pro‐
cess by which data governance is implemented are key to a successful implementation
of a data governance strategy.
The people and process are aspects of a data governance strategy that often get over‐
looked or oversimplified. There is an exceedingly heavier reliance on governance
tools, and though tools are getting better and more robust, they are not enough by
themselves; how the tools are implemented, an understanding of the people using
them, and the process set up for their proper use are all critical to governance success
as well.
The People: Roles, Responsibilities, and Hats
Many data governance frameworks revolve around a complex interplay of many roles
and responsibilities. These frameworks rely heavily on each role playing its part in
keeping the well-oiled data governance machine running smoothly.
57
The problem with this is that most companies are rarely able to exactly or even semifully match these frameworks, due to lack of employee skill set or, more commonly, a
simple lack of headcount. For this reason, employees working in the information and
data space of their company often wear different user hats. We use the term hat to
delineate the difference between an actual role or job title and tasks that are done. The
same person can perform tasks that align with many different roles or wear many dif‐
ferent hats as part of their day-to-day job.
User Hats Defined
In Chapter 1, we outlined three broad categories of governors, approvers, and users.
Here we will look in depth at the different hats (versus roles) within each category
(and expand on an additional category of ancillary hats), the tasks associated with
each hat, and finally, the implications and considerations when looking at these hats
from a task-oriented perspective versus a role-based approach. In Table 3-1 we’ve lis‐
ted out each hat with its respective category and key tasks for quick reference, with
more detailed descriptions following.
Table 3-1. Different hats with their respective categories and the tasks associated with them
Hat
Legal
Privacy tsar
Data owner
Data steward
Data analyst/data
scientist
Business analyst
Customer support
specialist
C-suite
External auditor
Category
Ancillary
Governor
Approver (can also be
governor)
Governor
User
Key tasks
Knows of and communicates legal requirements for compliance
Ensures compliance and oversees company’s governance strategy/process
Physically implements company’s governance strategy
(e.g., data architecture, tooling, data pipelining, etc.)
Performs categorization and classification of data
Runs complex data analytics/queries
User
Ancillary (can also be
a user)
Ancillary
Ancillary
Runs simple data analyses
Views customer data (but does not use this data for any kind of analytical
purpose)
Funds company’s governance strategy
Audits a company’s compliance with legal regulations
Legal (ancillary)
Contrary to the title of legal, this hat may or may not be an actual attorney. This hat
includes the tasks of ensuring the company is up to date in terms of compliance with
the legal requirements for its data handling and communicating this information
internally. Depending on the company, the person with this hat may actually be an
attorney (this is especially true for highly regulated companies, which will be covered
in depth later) who must have a deep knowledge of the type of data collected and how
it’s treated to ensure that the company is in compliance in the event of an external
audit. Other companies, especially those dealing with sensitive but not highly
58
|
Chapter 3: Ingredients of Data Governance: People and Processes
regulated data, are more likely to simply have someone whose task is to be up to date
on regulations and have a deep understanding of which ones apply to the data the
company collects.
Privacy tsar (governor)
We chose to title this hat privacy tsar because this is a term we use internally at Goo‐
gle, but this hat has also been called governance manager, director of privacy, and
director of data governance in other literature. The key tasks of this hat are those
which ensure that the regulations the legal department has deemed appropriate are
followed. Additionally, the privacy tsar also generally oversees the entire governance
process at the company, which includes defining which governance processes should
be followed and how. We will discuss other process approaches later in this chapter.
It’s important to note that the privacy tsar may or may not have an extremely techni‐
cal background. On the surface it might seem that this hat would come from a techni‐
cal background, but depending on the company and the resources it has dedicated to
its data governance efforts, these tasks are often performed by people who sit more
on the business side of the company rather than on the technical side.
Understanding the movement of people is of utmost importance when it comes to
battling COVID-19. Google, a company that processes significant amounts of highly
personal data that includes location information, was torn between helping health‐
care providers and authorities to battle the deadly pandemic more effectively and pre‐
serving the trust of the billions of people worldwide who use Google’s services.
Privacy tsar, work example 1: Community mobility reports. The challenge of preserving
privacy while at the same time providing useful, actionable data to health authorities
required the full attention of Google’s privacy tsars, the people entrusted with creating
the internal regulations that make sure that technology does not intrude into users’
personal data and privacy.
The solution they found was to provide information in an aggregated form, based on
anonymized sets of data from only those users who have turned on “location history”
in Google’s services.1 This setting is off by default, and users need to “opt in” to enable
it. Location information history can always be deleted by the user at any time. In
addition, differential privacy (a technique covered in Chapter 7) was further used to
identify small groups of users with outlier results and eliminate those groups com‐
pletely from the provided solution. Another differential privacy technique was
employed to add statistical noise to the results; the noise, statistically irrelevant for
aggregates, helps ensure that no individual can be tracked back through the data.
1 See Google’s “Community Mobility Reports” for more on this.
The People: Roles, Responsibilities, and Hats
|
59
The result is a useful set of reports tracking communities and changes in behavior
over time. Health officials can then assess whether “stay at home” orders are being
complied with, and trace back sources of infection due to people congregating.
In Figure 3-1, we see the results of that work. A sample from the report for San Fran‐
cisco County shows increased people presence in residential areas (bottom right) but
a reduced presence across the board in retail sites, grocery stores, parks, transit sta‐
tions, and workplaces. Note that no individual location is named, yet the data is use‐
ful for health officials in estimating where people congregate. For example, an order
about the reopening of retail stores can be considered.
Figure 3-1. Example from the Google mobility reports
Privacy tsar, work example 2: Exposure notifications. An even more daunting task related
to COVID-19 is how to safely (from a privacy perspective) inform people about pro‐
longed exposure to a person diagnosed with COVID-19.2 Because the virus is highly
contagious and can be transmitted through the air, identifying exposures and making
sure people who were inadvertently exposed to a positively diagnosed person get tes‐
ted (and isolate themselves if testing positive), is crucial to breaking infection chains
and limiting outbreaks. This process is a recognized technique in battling infections
2 For more about exposure notifications, see “Exposure Notifications: Using Technology to Help Public Health
Authorities Fight COVID‑19” and “Privacy: Preserving Contact Tracing”.
60
| Chapter 3: Ingredients of Data Governance: People and Processes
and is otherwise known as contact tracing. Technology can augment this technique by
immediately alerting the individual as an alternative to a prolonged phone investiga‐
tion in which public health authorities question a positively diagnosed individual as
to their whereabouts over the incubation period. (Many people cannot accurately
identify where they have been over the course of the past few days, nor can everyone
easily produce a list of all the people they have interacted with over the course of the
past week.)
However, a positive COVID-19 diagnosis is highly personal, and having this informa‐
tion delivered to everyone that a diagnosed person came in contact with is an emo‐
tionally loaded process. Furthermore, having your technology do that for you is
highly intrusive and will cause resistance to the point of not enabling this technology.
So how does a privacy tsar thread the needle between preserving personal informa‐
tion and privacy and combating a deadly disease?
The solution that was found maintains the principles required to ensure privacy:
• It must be an “opt-in” solution—the people must enable it, and the product pro‐
vides information to ensure consent is acquired after being informed.
• Since the topic is whether or not the subject was next to a diagnosed person,
location information, though it might be useful to health authorities to under‐
stand where the incident occurred, is not collected. This is a decision made by
the privacy tsar in favor of preserving privacy.
• The information is shared only with public health authorities and not with Goo‐
gle or Apple.
So how does the solution work? Every phone beams a unique yet random and fre‐
quently changing identifier to all nearby phones; the phones collect the list of bea‐
cons, and this list is matched with anyone who has reported their diagnosis. If your
phone was in proximity to the phone of someone who has uploaded a positive diag‐
nosis, the match will be reported to you (see Figure 3-2). Note that the identity of the
infected individual is not reported, nor is the specific time and place. Thus the crucial
information (you have been near a positively diagnosed individual so get tested!) is
shared without sacrificing the privacy of the infected individual.
The People: Roles, Responsibilities, and Hats
|
61
Figure 3-2. Excerpt from the Google/Apple guide to the exposure notification technology
Data owner (approver/governor)
In order for the privacy tsar’s governance strategy/process to be realized, the data
owner is needed.3 The tasks of the data owner include physically implementing the
3 While “classical literature” on data governance often separates data owners and data custodians (the former
residing more on the business side of things, and the latter more on the technical side), during the course of
our research and interviews with many companies we found that, in practice, these two “sides of the coin” are
often conflated, and the actual tasks of data ownership tend to fall on those with technical expertise.
62
| Chapter 3: Ingredients of Data Governance: People and Processes
processes and/or strategies laid out by the privacy tsar. This most often includes the
ideation and creation of the data architecture of the company, along with choosing
and implementing tooling and data pipeline and storage creation, monitoring, and
maintenance—in other words, “owning the data.” It’s clear from the task descriptions
that these must be performed by someone with quite a bit of technical background
and expertise; hence people who wear the data owner hat largely are engineers or
folks with an engineering background.
Data steward (governor)
When researching data governance, you will find that there is likely to be much
weight given to the role of data steward, and that’s with good reason—the tasks of a
data steward include categorization and classification of data. For any sort of gover‐
nance to be implemented, data must be defined and labeled to clearly identify what
the data is: sensitive, restricted, health related, and so on. A large part of the reason
we advocate the usage of the term hats versus roles is exemplified by the fact that it’s
very rare to find a singular person doing the “role” of data steward in the wild. The
act of data “stewardship” is highly, highly manual and extremely time consuming. In
fact, one company we spoke to said that for a short while it actually had a “full time”
data steward who quit after several months, citing the job as being “completely soul
sucking.” Because of the manual, time-consuming nature of the role—coupled with
the fact that in most cases there is no dedicated person to perform stewardship
duties—these duties often fall to many different people across the company or to
someone who has another role/other duties they perform as well. As such, full data
categorization/classification is often not done well, not done fully, or, in the worst
cases, just not done at all. This is an important item to note, because without steward‐
ship, governance is incomplete at best. We will cover this in more depth later, but
here is where we begin to see a recurring theme when looking at the people and the
process. Many of the governance processes that most companies employ right now
are undertaken to get around the fact that stewardship falls short. As we outlined in
Chapter 1, the quick growth and expansion of data collected by companies has resul‐
ted in an overwhelm of data, and companies simply don’t have the time or headcount
to dedicate to being able to categorize, classify, and label ALL of their data, so they
create and utilize other methods and strategies to “do their best given the limitations.”
Data analyst/data scientist (user)
Data analysts and data scientists are, in general, some of the primary or key users of
data within a company and are largely who data governance efforts are for. Compa‐
nies struggle with governance and security of their data versus the democratization of
their data. In order to be data driven, companies must collect and analyze large
amounts of data. The more data that can be made available to analysts and scientists,
the better—unless of course that data is sensitive and should have limited access.
The People: Roles, Responsibilities, and Hats
|
63
Therefore, the better the governance execution, the better (and the more safely) ana‐
lysts or scientists are able to do their job and provide valuable business insights.
Business analyst (user)
While data analysts and data scientists are the main users or consumers of data, there
are a few people in the periphery of a company who also use and view data. In mov‐
ing toward being more data driven, companies have folks who sit on the business side
of things who are very interested in the data analyses produced by analysts and scien‐
tists. In some companies, data engineers aid in the creation and maintenance of much
simpler analytics platforms to aid in “self-service” for business users. More and more
business people in companies have questions that they hope crunching some data will
help them to answer. Analysts/scientists thus end up fielding many, many inquiries,
some of which they simply don’t have time to answer. By enabling business users to
directly answer some of their own questions, analysts/scientists are able to free up
their time to answer more complex analysis questions.
Customer support specialists (user/ancillary)
While a customer support specialist is technically only a “viewer” of data, it’s worth
noting that there are folks with this role who will need access to some sensitive data,
even though they don’t have any tasks around manipulating that data. In terms of
hats, customer support specialists do tend to have this as their sole role and are not
also doing other things; however, they are consumers of data, and their needs, and
how to grant them appropriate access, must be considered and managed by the other
hats executing a company’s governance strategy.
C-suite (ancillary)
In many companies, members of the C-suite have limited tasks in relation to the
actual execution of a data governance strategy. They are nonetheless a critical hat in
the grand scheme of governance because they “hold the purse strings.” As we men‐
tioned earlier, tools and headcount are critical factors when considering a successful
data governance strategy. It thus makes sense that the person who actually funds that
strategy must understand it and be on board with the funding that makes it a reality.
External auditor (ancillary)
We have included the hat of external auditor in this section despite the fact that exter‐
nal auditors are not within a particular company. Many of the companies we’ve spo‐
ken to have mentioned the importance of the external auditor in their governance
strategy. No longer is it good enough to simply be “compliant” with regulations—
companies now often need to prove their compliance, which has direct implications
for the way a governance strategy and process is employed. Oftentimes, companies
need to prove who has access to what data as well as all the different locations and
64
|
Chapter 3: Ingredients of Data Governance: People and Processes
permutations of that data (its lineage). While internal tool reporting can generally
help with providing proof of compliance, the way that a governance strategy is set up
and tended to can help, or hinder, the production of this documentation.
Data Enrichment and Its Importance
As we mentioned at the beginning of this section, one person may wear many hats—
that is, perform many of the tasks involved in carrying out a data governance strategy
at a company. As can be seen in the list of hats and the (fairly long) list of tasks within
each hat shown in Table 3-1, it’s easy to see how many of these tasks may not be com‐
pleted well, if at all.
While there are many tasks that are important to successful implementation of a data
governance strategy, it could be argued that the most critical are data categorization,
classification, and labeling. As we pointed out, these tasks are manual and highly time
consuming, which means that they rarely get executed fully. There is a saying: “In
order to govern data, you must first know what it is.” Without proper data enrich‐
ment (the process of attaching metadata to data), the central task of the data steward
hat, proper data governance falls short. This central task of the data steward is so key,
however, that what we commonly see is that the person who wears the data steward
hat also often wears the privacy tsar hat as well as the data owner hat, and/or they
may even wear a hat in a completely different part of the company (a common one we
see is business analyst). Wearing many hats results in a limited amount of tasks that
can be done (one person can only do so much!), and most often the majority of the
time-consuming task of data enrichment falls off the list.
In the next section we will cover some common governance processes we’ve observed
over the years. One thing to keep in mind while reading through the processes is how
the hats play a role in the execution of these processes. We will discuss their interplay
later in this chapter.
The Process: Diverse Companies, Diverse Needs and
Approaches to Data Governance
It’s important to note that in the discussion of the people and processes around data
governance, there is no “one size fits all” approach. As mentioned, in this section we
will begin to examine some broad company categories, outline their specific con‐
cerns, needs, and considerations, and explore how those interplay with a company’s
approach to data governance.
The Process: Diverse Companies, Diverse Needs and Approaches to Data Governance
|
65
We’d like to reemphasize the point that there is no “one size fits all”
approach to governance; and you need to fit your program to your
needs—whether those be mitigated by headcount, the kind of data
you collect, your industry, etc. We have seen many governance
approaches and frameworks that are somewhat inflexible and thus
might be difficult to implement if you don’t fit their particular
parameters. We hope that through exploration of elements to con‐
sider you’re able to create a framework for yourself that not only
matches your governance needs and goals but also matches where
your company is right now, and where you’d like to be in the future.
Legacy
Legacy companies are defined as companies that have been around for quite some
time and most certainly have, or have had, legacy on-premises (on-prem) systems—
most often many different systems, which bring with them a host of issues. The time
and level of effort this work requires often leads to it not being done fully, or simply
not being done at all. Further, for consistency (and arguably for the most effective use
of data and data analytics), every company should have a central data dictionary,
defining all data names, classes, and categories that is standardized and used through‐
out the company. Many legacy companies lack this central data dictionary because
their data is spread out through these various on-prem systems. More often than not,
these on-prem systems and the data within them are associated with a particular
branch or line of business. And that branch, in and of itself, is agnostic of the data
that resides in the other lines of the business’s systems. As such, there ends up being a
dictionary for each system and line of business that may not align to any other line of
business or system, which makes cross-system governance and analytics nearly
impossible.
A prime example of this would be a large retailer that has a system that houses its
online sales and another system that handles its brick-and-mortar sales. In one sys‐
tem the income the company receives from a sale is called “revenue,” while in the
other system it’s simply called “sales.” Between the two systems, the same enterprise
dictionary is not being used—one can see where this becomes problematic if execu‐
tives are trying to figure out what the total income for the company is. Analytics are
nearly impossible to run when the same data does not have the same metadata ver‐
nacular attached. This becomes an even larger issue when considering sensitive data
and its treatment. If there is no agreed-upon, company-wide terminology (as outlined
in the enterprise dictionary), and if that terminology is not implemented for all sour‐
ces of data, governance is rendered incomplete and ineffective.
The power of the cloud and big data analytics drives many companies to want to
migrate their data, or a portion of their data, to the cloud—but the past pain of incon‐
sistent enterprise dictionaries and haphazard governance gives them pause. They
66
|
Chapter 3: Ingredients of Data Governance: People and Processes
don’t want to “repeat their past mistakes”; they don’t want to simply replicate all of
their current issues. While tools can help to right past wrongs, they simply aren’t
enough. Companies need (and desire) a framework for how to move their data and
have it properly governed from the beginning, with the right people working the right
process.
Cloud Native/Digital Only
Cloud-native companies (sometimes referred to as digital only) are defined as compa‐
nies who have, and have always had, all of their data stored in the cloud. Not always,
but in general, these tend to be much younger companies who have never had onpremises systems and thus have never had to “migrate” their data to a cloud environ‐
ment. Based on that alone, it’s easy to see why cloud-native companies don’t face the
same issues that legacy companies do.
Despite the fact that cloud-native companies do not have to deal with on-prem sys‐
tems and the “siloing” of data that often comes along with that, they still may deal
with different clouds as well as different storage solutions within and between those
clouds that have their own flavor of “siloing.” For one, a centralized data dictionary, as
we’ve explored, is already a challenge, and having one that spans multiple clouds is
even more daunting. Even if a centralized data dictionary is established, the process
and tools (because some clouds require that only certain tools be used) by which data
is enriched and governed in each cloud will probably be slightly different. And that is
something that hinders consistency (in both process and personnel) and thus hinders
effectiveness. Further, even within a cloud, there can be different storage solutions
that also may carry with them different structures of data (e.g., files versus tables ver‐
sus jpegs). These structures can be difficult to attach metadata to, which makes gover‐
nance additionally difficult.
In terms of cloud—and data governance within clouds specifically—cloud-native
companies tend to have better governance and data management processes set up
from the beginning. Because of their relatively young “age” in dealing with data, there
is often less data overall, and they have more familiarity with some common datahandling best practices. Additionally, cloud-native companies by definition have all
of their data in the cloud, including their most sensitive data—which means that
they’ve most likely been dealing with the need for governance since the beginning,
and they most likely have some processes in place for governing, if not all of their
data, at least their most sensitive data.
Retail
Retail companies are an interesting category, as not only do they often ingest quite a
bit of data from their own stores (or online stores), but they also tend to ingest and
utilize quite a bit of third-party data. This presents yet another instance in which data
The Process: Diverse Companies, Diverse Needs and Approaches to Data Governance
|
67
governance is only as good as the process set up for it and the people who are there to
execute it. The oft-mentioned importance of creating and implementing a data dictio‐
nary applies, as well as having a process around how this third-party data is ingested,
where it lands, and how governance can be applied.
One twist that we have not discussed but that very much applies in the case of retail is
that of governance beyond simple classification of data. So far we have discussed the
importance of classifying data (especially sensitive data) so that it can be known and
thus governed appropriately. That governance most often relates to the treatment of
and/or access to said data. But data access and data treatment are not the only aspects
to consider in some instances; the use case for that data is also important. In the case
of retail, there may be a certain class of data—email for example—that was collected
for the purpose of sending a customer their receipt from an in-store purchase. In the
context (use case) of accessing this data for the purposes of sending a customer a
receipt, this is completely acceptable. If, however, one wanted to access this same data
for the purpose of sending a customer some marketing material around the item they
just purchased, this use case would not be acceptable unless the customer has given
explicit consent for their email to be used for marketing. Now, depending on the
employee structure at a company, this problem may not be solvable with simple rolebased access controls (fulfillment gets access, marketing does not) if the same
employee may cover many different roles. This warrants the need for a more complex
process that includes establishing use cases for data classes.
Combo of Legacy and Retail
A particularly interesting use case we’ve come across in our customer interviews is
that of a very large legacy retail company. In business for over 75 years, this company
is looking to leverage powerful data analytics tools to help it move toward being a
more data-driven company.
The struggle, interestingly, is not only in its old legacy on-prem data-storage systems
but also in its internal processes around data and data management.
It currently has its data separated into several data marts, a configuration aimed at
distributing responsibility for segments of data: a mart for marketing for in-store
sales, for third-party sales, and so on. This, however, has become problematic for the
company, as not only is there no “central source of truth” in terms of an enterprise
dictionary, but it also cannot run any kind of analytics across its data marts, resulting
in duplication of data from one mart to another since analytics can only be run within
a mart.
Historically, the company was very focused on keeping all of its data on-prem; how‐
ever, this view has changed with the increased security that is now available in the
cloud, and the company is now looking to migrate all of its data off-premises.
Through this migration effort, not only is the company looking to change the basic
68
| Chapter 3: Ingredients of Data Governance: People and Processes
infrastructure of its data story from a decentralized one (multiple marts) to a central‐
ized one (centralized data warehouse), but it is also using this as an opportunity to
restructure its internal processes around data management and governance (see
Figure 3-3).
Figure 3-3. Breaking down data silos by restructuring internal processes is often a key
stage after enterprises consolidate their on-premises data into the cloud
Data centralization enables the company to have one enterprise dictionary that allows
for all new data—namely sensitive data—to be quickly marked and treated accord‐
ingly. Quick and easy handling of sensitive data also enables quicker and easier access
controls (as they need to be applied only once, as opposed to each mart). This not
only saves overhead effort but also allows for the implementation of more easy-to-use
self-service analytics tools for employees who may not be analysts by trade but, with
the right tools and access to data, are able to run simple analytics. This new process,
however, is really only good for new incoming data.
This company, like many legacy companies with a lot of historical data, is struggling
with how to handle the data it already has stored. It currently has 15 years of data
(~25 terabytes) stored on-premises, and while it knows there is a wealth of informa‐
tion there, the time and effort required to migrate, enrich, and curate all this data
seems daunting at best, especially when the payoff is not known.
Highly Regulated
Highly regulated companies represent the sector of companies that deal with
extremely sensitive data—data that often carries additional compliance requirements
beyond the usual handling of sensitive information. Some examples of highly
regulated companies would be those dealing with financial, pharmaceutical, or
healthcare services.
The Process: Diverse Companies, Diverse Needs and Approaches to Data Governance
|
69
Highly regulated companies deal with multiple kinds of sensitive data. They have to
juggle not only basic data governance best practices but also the additional regula‐
tions related to the data they collect and deal in, and they face regular audits to make
sure that they are aboveboard and compliant. As a result of this, many highly regula‐
ted companies are more sophisticated in their data-governance process. As they’ve
had to deal with compliance related to their data from the get-go, they often have bet‐
ter systems in place to identify and classify their sensitive data and treat it
appropriately.
Also, for many of these kinds of companies, their business is based solely around sen‐
sitive data, so not only do they tend to have better processes in place from the begin‐
ning, but also those processes most often include a more well-funded and
sophisticated organization of the people who handle that sensitive data. These com‐
panies tend to actually have people dedicated to each of the hats discussed earlier, and
that additional headcount, as we pointed out, can be the deciding factor in whether a
data-governance strategy is successful or unsuccessful.
One final note about highly regulated companies is that, as a result of the legal
requirements they face on the handling of certain data, they can function similarly to
legacy companies in that they have difficulty moving off-prem and/or trying out new
tools. Any tool that will touch sensitive data under a regulation must meet the stand‐
ards (fully) of that regulation. As an example, a tool or product in beta may be used
by a healthcare company only if it is HIPAA compliant. Many product betas, because
they’re used for early access and as a way to work out bugs, are not designed to meet
the most stringent compliance standards. While the final product may be compliant,
the beta generally is not, meaning that highly regulated companies often don’t get to
experiment with these new tools/products and thus have trouble migrating to new,
potentially better tools than the ones they’re currently using.
Unique Highly Regulated Organization: Hospital/University
When discussing highly regulated industries, finance and healthcare are the ones
most commonly referenced. In our discussions with customers we came across
another highly regulated industry that had some unique challenges: hospital/universi‐
ties. These “companies” are unique in that they collect a lot of clinical data from their
hospitals, but they also collect (and produce) a lot of research data through
university-sponsored research studies.
Each of these types of data comes with its own specific regulations and standards for
research—e.g., HIPAA covers clinical data, and the Institutional Review Board (IRB)
protects the rights and welfare of human research subjects recruited to participate in
research activities.
70
|
Chapter 3: Ingredients of Data Governance: People and Processes
One particular hospital/university we spoke to was looking at the use case of being
able to run secondary analytics across its clinical and research data. Currently it has
two main pipeline solutions for its data: one for clinical and one for research.
For its clinical pipelines, data is stored on-prem in a Clarity data warehouse for each
hospital, and only analysts with access to that database are able to run analytics.
For its research pipelines, each “lab” within the university has its own on-prem stor‐
age, and again, only analysts with access to that server can run analytics. It can be
seen that not only does this structure not allow for secondary analyses to be run
across labs, but they also can’t be run across clinical and research.
Knowing that there is a wealth of value in these secondary analytics, this hospital/
university decided to migrate a large portion of its clinical and research data to the
cloud, to get it all in one central location. In order to do this, however, much needs to
be done to make the data comply with healthcare and research regulations. As such,
the hospital/university created a specialized team dedicated to this migration effort;
that team’s role included tasks such as: creating an enterprise dictionary, enriching
data, reviewing the presence of sensitive data, reviewing policies attached to data,
applying new policies to data, and applying a standardized file structure.
Fortunately for this organization, it was able to secure funding for such an elaborate
undertaking, and yet it is still looking at ways to automate its process. Migration and
the setup of data in the cloud is but one hurdle—maintaining and managing a data
store going forward will require effort as well, and as such, tools that enable automa‐
tion are top of mind for this organization.
Small Companies
For our purposes, we define a small company as one that has fewer than one thou‐
sand employees. One of the benefits of small companies is their smaller employee
footprint.
Smaller companies often have small data-analytics teams, which means that there are
fewer people who actually need to touch data. This means that there is less risk over‐
all. One of the primary reasons to govern data is to make sure that sensitive data does
not fall into the wrong hands. Fewer employees also makes for a much less arduous
and less complicated process of setting up and maintaining access controls. As we
discussed earlier, access controls and policy management can get increasingly compli‐
cated, especially when factors such as use case for data come into play.
Another benefit of a small amount of people touching data is that there is often less
proliferation of datasets. Data analysts and scientists, as part of their jobs, create many
different views of datasets and joins (combining tables from different dataset sour‐
ces). This proliferation of data makes it hard to track where the data came from and
where it’s going to (not to mention who had and now has access). Fewer analysts/
The Process: Diverse Companies, Diverse Needs and Approaches to Data Governance
|
71
scientists mean fewer datasets/tables/joins, resulting in data that’s much easier to
track and thus easier to govern.
Large Companies
While there are many more company types, we will end our exploration of broad cat‐
egories with large companies, defined as companies with more than one thousand
employees.
If small companies have the governance benefit of dealing with less data, large com‐
panies deal with the reverse, in that they often deal with a lot of data. Large compa‐
nies not only generate a great deal of data themselves but also often deal with a lot of
third-party data. This results in immense difficulty in wrapping their arms around it
all; they are overwhelmed by the data and often struggle to govern even a portion of
it. As such, only some data gets enriched and curated, which means that only some of
their data is able to be used to drive insights.
Large companies often put processes in place to limit this overwhelm by choosing
only select data to enrich, curate, and thus govern. A common strategy for limiting
the data that a data steward must enrich is to select a number of categories and gov‐
ern only the data that falls within those categories. Another is to only govern known
pipelines of data (these are the pipelines consisting of the primary data a company
deals with, such as daily sales numbers from a retail store) and to handle “ad hoc”
pipeline data (engineers at times are asked to create new pipelines to address onetime or infrequent data analysis use cases) only as time allows or if absolutely
necessary.
It’s easy to see that these strategies (which we will discuss in more depth later) result
in a sort of iceberg of data where enriched (known), curated data sits at the top, and
below that is a mound of data that is, in the main, un-nriched and thus unknown.
Companies don’t know what this data is, which means they can’t govern it, and
ungoverned data is quite scary. It could be sensitive, it could be noncompliant, and
there could be dire consequences if it happens to get leaked. This causes much fear
and forces companies with this problem to use strategies to help mitigate their risk.
Not only do large companies deal with copious amounts of data, but they also tend to
have a much larger workforce of data analysts, data scientists, and data engineers—all
of whom need access to data to do their jobs. More data, plus more people who need
access to data, results in much more complicated (and often poorly managed) pro‐
cesses around access control. Access control is generally based on user role, which
should determine the data (and only that data) that they need access to. This strategy
may seem simple, but it is a difficult process to implement for two main reasons.
First, in order to know what data is critical to a particular user’s role, data must be
known. And we know from previous discussion in this book that the majority of data
a company has is unknown. This results in not only the inability to govern (it’s
72
|
Chapter 3: Ingredients of Data Governance: People and Processes
impossible to govern what you don’t know you have) but also the inability to know
what data is appropriate for whom. Data specialists still need to do their job, however,
so (too much) access is often granted at the expense of risk. Companies try to offset
this risk by creating a “culture of security,” which puts the onus on the employee to do
the right thing and not expose or misuse potentially sensitive information. Another
issue with role-based access is that roles and their ensuing uses for data are not always
black and white; the use case for data must also be considered. Depending on the
company, a user in the same role may be able to use data for some use cases and not
for others (an example being the use case we outlined in the retail section). As
touched on Chapter 2, and as will be covered in depth later, the addition of use case
as a parameter during policy creation helps to mitigate this issue.
Like legacy companies, large companies often also deal with the issue of having many
different storage systems, the majority of which are legacy. Large companies tend to
be built up over time, and with time comes more data and more storage systems.
We’ve already discussed how the siloed nature of different storage systems makes
governance difficult. Different storage systems naturally create data silos, but large
companies have another factor that results in even more complex “silos”: acquisitions.
Large companies are sometimes built completely from within, but others become
large (or larger) due to acquisitions. When companies acquire other, smaller compa‐
nies, they also acquire all their data (and its underlying storage), which brings along a
whole host of potential problems—primarily, how the acquired company handled its
data, as well as its overall governance process and approach. This includes how the
company managed its data: its method of data classification, its enterprise dictionary
(or lack thereof), its process and methods around access control, and its overall cul‐
ture of privacy and security. For these reasons, many large companies find it nearly
impossible to marry their central governance process with that of their acquisitions,
which often results in acquired data sitting in storage and not being used for
analytics.
People and Process Together: Considerations, Issues, and
Some Successful Strategies
We have now outlined the different people involved in various kinds of companies, as
well as some specific processes and approaches utilized by the different company
types.
There is obviously a synergy between people and process, of which there are some
considerations and issues. We will review several of the issues we’ve observed, along
with outlining a few strategies we have seen in which the right people and process
together have resulted in moving toward a successfully implemented data-governance
strategy.
People and Process Together: Considerations, Issues, and Some Successful Strategies
|
73
Considerations and Issues
This certainly is not an exhaustive list of all the potential issues with implementing a
successful data-governance strategy; we are simply highlighting some of the top
issues we’ve observed. We only briefly note the mitigation efforts to combat these
issues; the latter half of this text will go into these in much more detail.
“Hats” versus “roles” and company structure
We previously discussed our intentional use of the terms hats versus roles and the
impact that wearing many hats has on the data governance process. To expand on
that idea further, an additional issue that arises with hats versus roles is that responsi‐
bility and accountability become unclear. When looking at the different kinds of
approaches that different companies take to achieve governance, an underlying need
is for actual people to take responsibility for the parts of that process. This is easy
when it is clearly someone’s job to conduct a piece of the process, but when the lines
between what is and what is not within a person’s purview are blurred, these fuzzy
lines often result in inadequate work, miscommunication, and overall mismanage‐
ment. It’s clear that a successful governance strategy will rely not simply on roles but
on tasks, and on who is responsible or accountable for these tasks.
Tribal knowledge and subject matter experts (SMEs)
When talking with customers about their pain points with data governance, one of
the things we hear over and over again is that they need tools to help their analysts
find which datasets are “good,” so that when an analyst is searching for data, they
know that this dataset is of the best quality and is the most useful one for their use
case. The companies state that this would help their analysts save time searching for
the “right/best” dataset, as well as assisting them in producing better analytics. Cur‐
rently, through most companies, the way analysts know which datasets they should
work with is by word of mouth, or “tribal knowledge.” This is an obvious problem for
companies because roles change, people move on, etc. Companies request “crowd‐
sourcing” functionality, such as allowing analysts to comment on or rank datasets to
help give them a “usefulness” score for others to see when searching. This suggestion
is not without merit but uncovers the larger problems of findability and quality.
Companies are relying on people to know the usefulness of a dataset and to transfer
that knowledge on to others, yet this strategy is fallible and difficult (if not impossi‐
ble) to scale. This is where tools that lessen (or negate) the effort placed on a particu‐
lar user or users aid in the process. A tool that can detectthe most-used datasets and
surface these first in a search, for example, can help minimize the reliance on tribal
knowledge and SMEs.
74
|
Chapter 3: Ingredients of Data Governance: People and Processes
Definition of data
Regardless of their type, all companies want to be able to collect data that can be used
to drive informed business decisions. They are certain that more data, and the analyt‐
ics that could be run on that data, could result in key insights—insights that have the
potential to skyrocket the success of their business. The problem, however, is that in
order for data to be used, it must be known. It must be known what the letters or
numbers or characters in a column of a table mean. And now it must also be known
whether those numbers, letters, or characters represent information that is sensitive
in nature and thus needs to be treated in a specific way. Data enrichment is key to
“knowing” data, yet it is largely a manual process. It generally requires actual people
to look at each and every piece of data to determine what it is. As we discussed, this
process is cumbersome on its own, and it becomes almost impossible when the extra
complexity of disparate data storage systems and different data definitions and cata‐
logs are considered. The “impossible” nature of this work in general means that it just
never gets done; this leaves companies scrambling to use a few tools and to imple‐
ment some half strategies to make up for it—and also hoping that educating people
on how data should be treated and handled will somehow be enough.
Old access methods
Gone are the days of simple access controls—i.e., these users/roles get access and
these users/roles do not. Historically, there were not many users or roles who even
had the need to view or interact with data, which meant that only a small portion of
employees in a company needed to be granted access in the first place. In today’s
data-driven businesses there is the potential for many, many users who may need to
touch data in a variety of ways, as seen in the beginning of this chapter. Each hat has
different types of tasks that it may need to do in relation to data that requires varying
levels of access and security privilege.
We already discussed the problematic nature of unknown data; another layer to that
is the implementation of access controls. There is a synergy between knowing which
data even needs access restrictions (remember, data must be known in order for it to
be governed) and knowing what those restrictions should be for what users. As dis‐
cussed in Chapter 2, there are varying levels of access control, all the way from access
to plain text to access to hashed or aggregated data.
A further complication around access is that of the intent of the user accessing data.
There may be use cases for which access can and should be granted, and other use
cases for which access should strictly be denied. A prime example (and one we’ve
heard from more than one company) is a customer’s shipping address. Imagine the
customer just bought a new couch, and their address was recorded in order to fulfill
shipment of that couch. A user working to fulfill shipping orders should undoubtedly
get access to this information. Now let’s say that the slipcovers for the couch this cus‐
tomer just bought go on sale, and the company would like to send a marketing flyer
People and Process Together: Considerations, Issues, and Some Successful Strategies
|
75
to the customer to let them know about this sale. It may be the case that the user in
the company who handles shipping data also happens to handle marketing data
(remember hats?). If the customer has opted OUT of promotional mail, the user in
this example would not be able to send the marketing material even though they have
access to that data. This means that access controls and policies need to be sensitive
enough to account not only for a black-and-white “gets access/doesn’t get access” rule
for a particular user, but also for what purpose the user is using the data.
Regulation compliance
Another struggle that companies have is around compliance with regulations. Some
regulations, such as those in financial and healthcare industries, have been around for
quite a while, and as we pointed out before, companies who deal with these kinds of
data tend to have better governance strategies, for two main reasons: one, they’ve
been dealing with these regulations for some time and have built in processes to
address them, and two, these regulations are fairly established and don’t change
much.
The advent of a proliferation of data collection has brought about new regulations
such as GDPR and CCPA that aim to protect all of a person’s data, not just their most
sensitive data (i.e., health or financial data). Companies in all types of industries, not
just highly regulated ones, now must comply with these new regulations or face seri‐
ous financial consequences. This is a difficult endeavor for companies that previously
did not have regulations to comply with and thus perhaps did not address this during
the setting up of their data infrastructure. As an example, one of the main compo‐
nents of GDPR is the “right to be forgotten,” or the ability for a person to request that
all their data collected by a company be deleted. If the company does not have its data
set up to find all the permutations of a person’s individual data, it will struggle to be
compliant. Thus, it can be seen how findability is important not only from the per‐
spective of finding the right data to analyze but also from the perspective of finding
the right data to delete.
Case Study: Gaming of Metrics and Its Effect on Data Governance
Be aware of the human element when you assign regulations and compliance. A rele‐
vant case study about how, when introducing metrics, you should factor in the human
response to these metrics is in Washington, DC’s school system. In 2009, Washington,
DC introduced a new ranking system called IMPACT through which teachers were
assessed and scored. The system, born out of the intention to promote “good” teach‐
ers and generally improve education—and, even further, to allow “low scoring” teach‐
ers to be dismissed—actually did not achieve the desired result.
The ranking system was based on an algorithm that took into account standardized
test scores of students, with the underlying assumption that test scores for students
76
|
Chapter 3: Ingredients of Data Governance: People and Processes
with “good” teachers should show improvement year after year. However, the ranking
system did not include social and familial circumstances, nor did it include any kind
of feedback or training from a controlled set.
The result of implementing the system was that teachers were let go based only on the
results of the standardized tests, and without regard to feedback from administrators.
The immediate reaction was that teachers pivoted to focusing on how to pass the
standardized tests rather than on the underlying principles of the subjects. This resul‐
ted in high scores and advancement for teachers whose students passed the tests, but
it did not, sadly, promote excellence among the students.
In fact, in 2018 a city-commissioned investigation revealed that one in three students
graduating in 2017 received a diploma despite missing classes. After being challenged,
and after a detailed account of the usage of the algorithm and its usefulness had been
carried out, the system was eventually “reconsidered” in 2019.4
The process lessons here are that, although the motivation (improving teaching, pro‐
moting excellent teachers) should have been easily agreed upon by both teachers’
unions and the school district administrators, the implementation was lacking. A
proper way to introduce a process is to first have a discussion with the affected peo‐
ple. Take feedback to heart and act on it. Run the process for the first time as a trial,
with transparent results, open discussion, and a willingness to pivot if it isn’t working
as intended.
Processes and Strategies with Varying Success
The issues with people and process are many, and we have observed some strategies
that have been implemented with varying degrees of success. While we will explore
some of these strategies here, the latter part of this book dives deeper into how these
strategies (and others) can be implemented to achieve greater data governance
success.
Data segregation within storage systems
We’ve reviewed some of the issues that arise from having multiple data-storage sys‐
tems; however, some companies use multiple storage systems and even different stor‐
age “areas” within the same storage system in an advantageous and systematic way.
Their process is to separate curated/known data from uncurated/unknown data. They
may do this in a variety of ways, but we have heard two common, prevailing
strategies.
4 Perry Stein, “Chancellor Pledges to Review D.C.’s Controversial Teacher Evaluation System”, Washington Post,
October 20, 2019.
People and Process Together: Considerations, Issues, and Some Successful Strategies
|
77
The first strategy is to keep all uncurated data in an on-prem storage system and to
push curated data that can be used for analytics to the cloud. The benefit that compa‐
nies see in this strategy is that the “blast radius,” or the potential for data to be leaked
either by mistake or by a bad actor, is greatly diminished if only known, clean, and
curated data is moved into a public cloud. To be clear, companies often indicate that it
is not their total distrust of cloud security that is the problem (although that can play
a role as well); it is their concern that their own employees may unwittingly leak data if
it’s in a public cloud versus an on-prem environment.
Figure 3-4 shows an example of data stored on-premises and in the cloud. As you can
see, even though both storage systems have structured datasets, only the cloud dataset
has been enriched with metadata and thus has been treated with the appropriate gov‐
ernance controls (hashing in this case). The on-premises dataset, while it has the
same information, hasn’t been curated or enriched; we don’t know just by looking at it
that the numbers in the first column are credit card numbers or that the words in the
second column are customer names. Once these columns have been curated and the
appropriate metadata (credit card number, customer name, etc.) has been attached,
we can attach governance controls, so that even if the data gets leaked, it will be the
protected version and not plain text.
Figure 3-4. Enriched versus not enriched datasets on-premises and in the public cloud
As we’ve stated, an obvious benefit of this strategy is that if sensitive data resides only
on-premises, then if it’s leaked or incorrectly accessed, that could only have been
done by someone within the company. On the other hand, if that same sensitive data
is in a public cloud, it could potentially have been accessed by anyone if leaked or
hacked into.
78
|
Chapter 3: Ingredients of Data Governance: People and Processes
There are a few drawbacks to this strategy, however. One is that when data is segrega‐
ted this way—some residing on premise and some residing in the cloud—crossstorage analytics are difficult if not impossible to complete. Since one of the main
drivers of collecting so much data is being able to run powerful analytics, hindering
this seems counterproductive. The other drawback is that segregation also requires
upkeep of and attendance to these multiple storage systems and data pipelines, not to
mention the creation, maintenance, and enforcement of additional access controls, all
of which are difficult to properly manage and stay on top of over time.
The second strategy is similar to the first in that there is a separation between curated
and uncurated data, but this is done within the same cloud environment. Companies
will create different layers or zones (as can be seen in Table 3-2) within their cloud
environment and base access on these; the bottom, uncurated zone may be accessed
by only a few users, while the uppermost zone, curated and cleaned (of sensitive
data), may be accessed by any user.
Table 3-2. A “data lake” within cloud storage with multiple tiers showing what kind of data
resides in each and who has access
Insights
zone
Staging
zone
Raw zone
Types of data
Known, enriched, curated, and cleaned data. Data also has likely had
governance controls such as encryption, hashing, redaction, etc.
Example: Well-labeled, structured datasets.
More known and structured data. Data from multiple sources is likely
to be joined here. This is also where data engineers prep data, cleanse
it, and get it ready to drop into the insights zone.
Any kind of data. Unstructured and uncurated. Could also include
things such as videos, text files, etc. Example: Video files, unstructured
datasets
Access
Highest level of access. Most if not all
data analysts/scientists and others in a
user role.
More access. Mostly data engineers—
those in charge of putting together
datasets for analytics.
Very restricted access. Likely a handful
of people or just an admin.
Clearly the benefits and drawbacks of this strategy are nearly a reverse of those of the
previous strategy. In this strategy, management and upkeep of the storage system,
data pipelines, and policies are limited to one system, which makes it far simpler and
more streamlined to stay on top of. Analytics are also largely easier to run because
they can all be run within one central storage system, as opposed to trying to run
them across multiple systems, or dealing with the constant moving of data from one
storage system to another for the purposes of analysis.
As we’ve noted, all data residing within a public cloud does carry the potential draw‐
back of data leaking (whether intentionally or unintentionally) out onto the public
internet—a cause for trepidation to be sure.
People and Process Together: Considerations, Issues, and Some Successful Strategies
|
79
Data segregation and ownership by line of business
As we’ve mentioned more than a few times, data enrichment is a key challenge in suc‐
cessful data governance for many reasons, the primary ones being level of effort and
lack of accountability/ownership.
One way we’ve observed companies handle this issue is by segregating their data by
line of business. In this strategy, each line of business has dedicated people to do the
work of governance on just that data. While this is often not actually delineated by
role, each line of business has a deep knowledge of the kind of business it has and
handles the ingress and egress of data (pipelines), data enrichment, enforcement/
management of access controls and governance policies, and data analysis. Depend‐
ing on the company size and/or complexities of data within each line of business,
these tasks may be handled by just one person, a handful of people, or a large team.
There are several reasons this process tends to be quite successful. The first is that the
amount of data any given “team” must wrap its arms around is smaller. Instead of
having to look at and manage a company’s entire data story, the team needs to under‐
stand and work on only a portion of it. Not only does this result in less work, but it
also allows for deeper knowledge of that data. Deeper knowledge of data has a multi‐
tude of advantages, a few being quicker data enrichment and the ability to run
quicker, more robust analytics.
The second reason this process is successful is that there is clear, identifiable owner‐
ship and accountability for data: there is a specific person (or persons) to go to when
something goes wrong or when something needs to change (such as the addition of a
new data source or the implementation of a new compliance policy). When there is
no clear accountability and responsibility for data, it’s easy for that data to be lost or
forgotten or worse—mismanaged.
In Figure 3-5, we’ve tried to show an example flow of different lines of business into a
central repository. As you can see, retail store sales, marketing, online sales, and HR
not only all feed into the central enterprise data repository in this example, but they
are also fed by this repository.
To give you an idea of the different key hats and their tasks in each line of business,
let’s take marketing as an example. In this line of business in our example, we have
the hats of data owner, data steward, and business analyst.
The data owner sets up and manages the pipelines in this line of business, along with
managing requests for new pipelines and ingestion sources. They also perform the
tasks of monitoring, troubleshooting, and fixing any data quality issues that arise, as
well as implementing any of the technical aspects of the company’s governance poli‐
cies and strategies.
80
|
Chapter 3: Ingredients of Data Governance: People and Processes
The data steward is the subject matter expert (SME) in this line of business; knowing
what data resides here, what it means, how it should be categorized/classed, and what
data is sensitive and what isn’t. They also serve as the point of contact between their
line of business and the central governing body at their company for staying up to
date on compliance and regulations, and they’re responsible for ensuring that the data
in their line of business is in compliance.
Finally, the business analyst is the expert on the business implications for the data in
this line of business. They’re also responsible for knowing how their data fits into the
broader enterprise, and for communicating which data from their line of business
should be used in enterprise analytics. In addition, they need to know what addi‐
tional/new data will need to be collected for this particular line of business to help
answer whatever the current or future business questions are.
From this example you can see how each hat has their role and tasks for just this one
line of business, and how a breakdown of any of those tasks can result in a less than
efficient flow/implementation of a governance program.
Figure 3-5. Flowchart of an enterprise with four lines of business and their data ingress/
egress, as well as a close-up of one line of business, its key hats
This process, while successful, does come with some pitfalls, however. The primary
one is that segregating data by line of business encourages the siloing of data and can,
depending on how it’s set up, inhibit cross-company analytics.
We recently worked with a large retail company that was dealing with this exact issue.
People and Process Together: Considerations, Issues, and Some Successful Strategies
|
81
This large US retailer (with over 350,000 employees), in an effort to better deal with
the tremendous amount of data it collects, divided up its data by several lines of busi‐
ness. These included air shipping, ground shipping, retail store sales, and marketing.
Having the data separated in this way greatly enabled the company to devote specific
data analytics teams to each line of business, whereby it could more quickly enrich its
data, allowing it to apply governance policies to aid with CCPA compliance. This
strategy, however, created issues when the company began to want to run analytics
across its different lines of business. To separate its data, the company created infra‐
structure and storage solutions with data pipelines that fed directly (and only) into
one or another line of business. We won’t go into too much depth on data pipelines
here, but in short, it is a common practice to have data “land” in only one storage
solution, because duplicating data and transferring it to additional storage areas is
costly and difficult to maintain. Because specific data resided only in one storage
area/silo, the company could not run analytics across its lines of business to see what
patterns might emerge between, for example, air shipping and retail store sales
(Figure 3-6).
Figure 3-6. The preceding example company’s two data silos: air shipping and ground
shipping. Each silo has its own data pipeline and stores this data within its own bucket,
meaning analytics can be run only within silos unless data from that pipeline is duplica‐
ted and piped into another silo.
This process of segregation by line of business to aid with accountability and respon‐
sibility is not a bad strategy, and while it’s obvious there are pitfalls, there are also
ways to make it more successful, which we will cover later in the book.
Creation of “views” of datasets
A classic strategy employed by many companies is to create different “views” of
datasets. As seen in Table 3-3, these views are really just different versions of the same
dataset and/or table that have sensitive information sanitized or removed.
82
|
Chapter 3: Ingredients of Data Governance: People and Processes
Table 3-3. Three potentially different types of “views” of data: one plain text, one with
sensitive data hashed, and another with sensitive data redacted
Plain text customer name
Anderson, Dan
Buchanan, Cynthia
Drexel, Frieda
Harris, Javiar
Hashed customer name
Anderson, #####
Buchanan, #####
Drexel, #####
Harris, #####
Redacted customer name
********
********
********
********
This strategy is classic (and works) because it allows analytics to easily be run, worry
free, by virtually anyone on the “clean” view (the one with sensitive data either hashed
or removed). It takes away the risk of access to and usage of sensitive data.
While this strategy works, it is problematic in the long run for several reasons. The
first is that it takes quite a bit of effort and manpower to create these views. The clean
view has to be manually created by someone who does have access to all of the data.
They must go in and identify any and all columns, rows, or cells that contain sensitive
data and decide how they should be treated: hashed, aggregated, completely removed,
etc. They then must create an entirely new dataset/table with these treatments in place
and make it available to be used for analytics.
The second issue is that new views constantly need to be created as fresher data
comes in. This results not only in much time and effort being put into creating “fresh”
views but also in a proliferation of datasets/tables that are difficult to manage. Once
all these views are created (and re-created), it’s hard to know which is the most fresh
and should be used. Past datasets/tables often can’t be immediately deprecated, as
there may be a need down the line for that specific data.
While views have some success and merit in aiding in access controls, we will pro‐
pose and discuss some strategies we have seen that not only are easier to manage but
scale better as a company collects more and more data.
A culture of privacy and security
The final process we’d like to discuss is that of creating a culture of privacy and secu‐
rity. While certainly every company and employee should respect data privacy and
security, the ways in which this strategy is thought through and implemented are
truly a special ingredient in creating not just a good data governance strategy but a
successful one.
We have dedicated an entire chapter to the strategy of building a successful data cul‐
ture (see Chapter 9), so more on that is to come; however, as a short introduction to
the concept, we will note that we have seen companies driven to really look at their
data culture and work to implement a new (successful) one because of one (or all) of
the following occurring within their organization:
People and Process Together: Considerations, Issues, and Some Successful Strategies
|
83
• Their governance/data management tools are not working sufficiently (meaning
the tools do not in and of themselves provide all of the governance “functional‐
ity” desired by a company).
• People are accessing data they should not be accessing and/or using data in ways
that are not appropriate (whether intentionally or not).
• People are not following processes and procedures put into place (again, whether
intentionally or not).
• The company knows that governance standards/data compliance are not being
met and don’t know what else to do, but they hope that educating people on
doing the “right thing” will help.
To be sure, throughout this text we have already covered in great detail governance
tools, the people involved, and the processes that are/can be followed. This is where
the importance of putting all the pieces together can be seen—i.e., tools, people, and
process. One or even two pieces are insufficient to achieve a successful governance
strategy.
As evidenced by the four reasons we’ve just listed even though a few tools, some peo‐
ple, and several processes are in place, some gaps may still remain.
This is where it must all come together—in a collective data culture that encompasses
and embodies how the company thinks about and will execute its their governance
strategy. That encompasses what tools it will use, what people it will need, and the
exact processes that will bring it all together.
Summary
In this chapter, we have reviewed multiple unique considerations regarding the peo‐
ple and process of data governance for different kinds of companies. We’ve also cov‐
ered some issues commonly faced by companies, as well as some strategies we’ve seen
implemented with varying success.
From our discussion it should be clear that data governance is not simply an imple‐
mentation of tools but that the overall process and consideration of the people
involved—while it may vary slightly from company to company or industry to indus‐
try—is important and necessary for a successful data governance program.
The process for how to think about data, how it should be handled and classified
from the beginning, how it continually needs to be (re)classified and (re)categorized,
and who will do this work and be responsible for it—coupled with the tools that
enable these tasks to be done efficiently and effectively—is key, if not mandatory, for
successful data governance.
84
|
Chapter 3: Ingredients of Data Governance: People and Processes
CHAPTER 4
Data Governance over a Data Life Cycle
In previous chapters, we introduced governance, what it means, and the tools and
processes that make governance a reality, as well as the people and process aspects of
governance. This chapter will bring together those concepts and provide a data life
cycle approach to operationalize governance within your organization.
You will learn about a data life cycle, the different phases of a data life cycle, data life
cycle management, applying data governance over a data life cycle, crafting a data
governance policy, best practices along each life cycle phase, applicable examples, and
considerations for implementing governance. For some, this chapter will validate
what you already know; for others, it will help you ponder, plant seeds, and consider
how these learnings can be applied within your organization. This chapter will intro‐
duce and address a lot of concepts that will help you get started on the journey to
making governance a reality. Before getting into the detailed aspects of governance,
it’s important to center our understanding on data life cycle management and what it
means for governance.
What Is a Data Life Cycle?
Defining what a data life cycle is should be easy—but in reality, it’s quite complex. If
you look up the definition of a data life cycle and its phases, you quickly realize that it
varies from one author to another, and from one organization to another. There’s
honestly not one right way to think about the different stages a piece of data goes
through; however, we can all agree that each phase that is defined has certain charac‐
teristics that are important in distinguishing it from the other phases. And because of
these different characteristics within each phase, the way to think about governance
will also vary as each piece of data moves through the data life cycle. In this chapter,
we will define a data life cycle as the order of stages a piece of data goes through from
85
its initial generation or capture to its eventual archival or deletion at the end of its
useful life.
It’s important to quickly point out that this definition tries to capture the essence of
what happens to a piece of data; however, not all data goes through each phase, and
these phases are simply logical dependencies and not actual data flows.
Organizations work with transactional data as well as with analytical data. In this
chapter, we will primarily focus on the analytics data life cycle, from the point when
data is ingested into a platform all the way to when it is analyzed, visualized, purged,
and archived.
Transactional systems are databases that are optimized to run day-to-day transactional
operations. These are fully optimized systems that allow for a high number of concur‐
rent users and transaction types. Even though these systems generate data, most are
not optimized to run analytics processes. On the other hand, analytical systems are
optimized to run analytical processes. These databases store historical data from vari‐
ous sources, including CRM, IOT sensors, logs, transactional data (sales, inventory),
and many more. These systems allow data analysts, business analysts, and even exec‐
utives to run queries and reports against the data stored in the analytic database.
As you can quickly see, transactional data and analytical data can have completely
different data life cycles depending on what an organization chooses to do. That said,
for many organizations, transactional data is usually moved to an analytics system for
analysis and will therefore undergo the phases of a data life cycle that we will outline
in the following section.
Proper oversight of data throughout its life cycle is essential for optimizing its useful‐
ness and minimizing the potential for errors. Data governance is at the core of mak‐
ing data work for businesses. Defining this process end-to-end across the data life
cycle is needed to operationalize data governance and make it a reality. And because
each phase has distinct governance needs, this ultimately helps the mission of data
governance.
Phases of a Data Life Cycle
As mentioned earlier, you will see a data life cycle represented in many different ways,
and there’s no right or wrong answer. Whichever framework you choose to use for
your organization has to be the one guiding the processes and procedures you put in
place. Each phase of the data life cycle as shown in Figure 4-1 has distinct characteris‐
tics. In this section, we will go through each phase of the life cycle as we define it,
delve into what each phase means, and walk through the implications for each phase
as you think about governance.
86
| Chapter 4: Data Governance over a Data Life Cycle
Figure 4-1. Phases of a data life cycle
Data Creation
The first phase of the data life cycle is the creation or capture of data. Data is gener‐
ated from multiple sources, in different formats such as structured or unstructured
data, and in different frequencies (batch or stream). Customers can choose to use
existing data connectors, build ETL pipelines, and/or leverage third-party ingestion
tools to load data into a data platform or storage system. Metadata—data about
data—can also be created and captured in this phase. You will notice data creation
and data capture used interchangeably, mostly because of the source of data. When
new data is created, that is referred to as data creation, and when existing data is fun‐
neled into a system, it is referred to as data capture.
In Chapter 1, we mentioned that the rate at which data is generated is growing expo‐
nentially, with IDC predicting that worldwide data will grow to 175 zettabytes by
2025.1 This is enormous! Data is typically created in one of these three ways:
Data acquisition
When an organization acquires data that has been produced by a third-party
organization
Data entry
When new data is manually entered by humans or devices within the
organization
Data capture
When data generated by various devices in an organization, such as IOT sensors,
is captured
It’s important to mention that each of these ways of generating data offers significant
data governance challenges. For example, what are the different checks and balances
for data acquired from outside your organization? There are probably contracts and
agreements that outline how the enterprise is allowed to use this data and for what
purposes. There might also be limitations as to who can access that specific data. All
these offer considerations and implications for governance. Later in the chapter, we
will look at how to think about governance during this phase, and we will call out the
different tools you should think about when designing your governance strategy.
1 Andy Patrizio, “IDC: Expect 175 Zettabytes of Data Worldwide by 2025”, Network World, December 3, 2018.
Phases of a Data Life Cycle
|
87
Data Processing
Once data has been captured, it is then processed, without yet deriving any value
from it for the enterprise. This is done prior to its use. Data processing is also referred
to as data maintenance, and this is when data goes through processes such as integra‐
tion, cleaning, scrubbing, or extract-transform-load (ETL) to get it ready for storage
and eventual analysis.
In this phase, some of the governance implications that you will come across are data
lineage, data quality, and data classification. All these have been discussed in much
more detail in Chapter 2. To make governance a reality, how do you make sure that as
data is being processed, its lineage is tracked and maintained? In addition, checking
data quality is very important to make sure you’re not missing any important values
before storing this data. You should also think about data classification. How are you
dealing with sensitive information? What is it? How are you ensuring management of
and access to this data so it doesn’t get into the wrong hands? Finally, as this data is
moving, it needs to be encrypted in transit and then later at rest. There are a lot of
governance considerations during this phase. We will delve into these concepts later
in the chapter.
Data Storage
The third phase in the data life cycle is data storage, where both data and metadata are
stored on storage systems and devices with the appropriate levels of protection.
Because we’re focusing on the analytics data life cycle, a storage system could be a
data warehouse, a data mart, or a data lake. Data should be encrypted at rest to pro‐
tect it from intrusions and attacks. In addition, data needs to be backed up to ensure
redundancy in the event of a data loss, accidental deletion, or disaster.
Data Usage
The data usage phase is important to understanding how data is consumed within an
organization to support the organization’s objectives and operations. In this phase,
data becomes truly useful and empowers the organization to make informed business
decisions when it can be viewed, analyzed, and/or visualized for insights. In this
phase, users get to ask all types of questions of the data, via a user interface or busi‐
ness intelligence tools, with the hope of getting “good” answers. This is where the
rubber meets the road, especially when confirming whether the governance processes
already instituted in previous phases truly work. If data quality is not implemented
correctly, the types of answers you receive will be incorrect or might not make too
much sense, and this could potentially jeopardize your business operations.
88
|
Chapter 4: Data Governance over a Data Life Cycle
In this phase, data itself may be the product or service that the organization offers. If
data is indeed the product, then different governance policies need to be enacted to
ensure proper handling of this data.
Because data is consumed by multiple internal and external stakeholders and pro‐
cesses during this phase, proper access management and audits are key. In addition,
there might be regulatory or contractual constraints on how data may actually be
used, and part of the role of data governance is to ensure that these constraints are
observed accordingly.
Data Archiving
In the data archiving phase, data is removed from all active production environments
and copied to another environment. It is no longer processed, used, or published, but
is stored in case it is needed again in an active production environment. Because the
volume of data generated is growing, it’s inevitable that the volume of archived data
grows. In this phase, no maintenance or general usage occurs. A data governance
plan should guide the retention of this data and define the length of time it will be
stored, including the different controls that will be applied to this data.
Data Destruction
In this final phase, data is destroyed. Data destruction, or purging, refers to the
removal of every copy of data from an organization, typically done from an archive
storage location. Even if you wanted to save all your data forever, it’s just not feasible.
It’s very expensive to store data that is not in use, and compliance issues create the
need to get rid of data you no longer need. The primary challenge of this phase is
ensuring that all the data is properly destroyed and at the right time.
Before destroying any data, it is critical to confirm whether there are any policies in
place that would require you to retain the data for a certain period of time. Coming
up with the right timeline for this cycle means understanding state and federal regu‐
lations, industry standards, and governance policies to ensure that the right steps are
taken. You will also need to prove that the purge has been done properly, which
ensures that data doesn’t consume more resources than necessary at the end of its
useful life.
You should now have a solid understanding about the different phases of a data life
cycle and what some of the governance implications are. As stated previously, these
phases are logical dependencies and not necessarily actual data flows. Some pieces of
data might go back and forth between different processing systems before being
stored. And some that are stored in a data lake might skip processing altogether and
get stored first, and then get processed later. Data does not need to pass through all
the phases.
Phases of a Data Life Cycle
|
89
We’re sure you’ve heard the phrase “Rome was not built in a day,” but that’s really
what this data life cycle is trying to do. Applying data governance in an organization
is a daunting task and can be very overwhelming. However, if you think about your
data within these logical data life cycle phases, implementing governance can be a
task that can be broken down into each phase and then thought through and imple‐
mented accordingly.
Data Life Cycle Management
Now that you understand the data life cycle, another common term you will run into
is data life cycle management (DLM). What’s interesting is that many authors will
refer to data life cycle and data life cycle management interchangeably. Even though
there might be a need or desire to bundle them together, it’s important to realize that
a data life cycle can exist without data life cycle management. DLM, therefore, refers
to a comprehensive policy-based approach to manage the flow of data throughout its
life cycle, from creation to the time when it becomes obsolete and is purged. When an
organization is able to define and organize the life cycle processes and practices into
repeatable steps for the company, this refers to DLM. As you start learning about
DLM, you will quickly run into a data management plan. So let’s quickly look at what
that means and what it entails.
Data Management Plan
A data management plan (DMP) defines how data will be managed, described, and
stored. In addition, it defines standards you will use and how data will be handled
and protected throughout its life cycle. You will primarily see data management plans
required to drive research projects within institutions, but the concepts of the process
are fundamental to implementing governance. Because of this, it’s worth us doing a
deep dive into them and seeing how these could be applied to implement governance
within an organization.
With governance, you will quickly realize that there is no a lack of templates and
frameworks—see, for example, the DMPTool from Massachusetts Institute of Tech‐
nology. You simply need to pick a plan or framework that works for your project and
organization and march ahead; there’s not one right or wrong way to do it. If you
choose to use a data management plan, here is some quick guidance to get you
started. The concepts here are much more fundamental than the template or frame‐
work, so if you were able to capture these in a document, then you’d be ahead of the
curve.
90
| Chapter 4: Data Governance over a Data Life Cycle
Guidance 1: Identify the data to be captured or collected
Data volume is important to helping you determine infrastructure costs and people
time. You need to know how much data you’re expecting and the types of data you
will be collecting:
Types
Outline the various types of data you will be collecting. Are they structured or
unstructured? This will help determine the right infrastructure to use.
Sources
Where is the data coming from? Are there restrictions on how this data can be
used or manipulated? What are those rules? All of these need to be documented.
Volume
This can be a little difficult to predict, especially with the exponential growth in
data; however, planning for that increase early on and projecting what it could be
would set you apart and help you be prepared for the future.
Guidance 2: Define how the data will be organized
Now that you know the type, sources, and volume of data you’re collecting, you need
to determine how that data will be managed. What tools do you need across the data
life cycle? Do you need a data warehouse? Which type, and from which vendor? Or
do you need a data lake? Or do you need both? Understanding these implications and
what each means will allow you to better define what your governance policies should
be. There are many regulations that govern how data can and cannot be used, and
understanding them is vital.
Guidance 3: Document a data storage and preservation strategy
Disasters happen, and ensuring that you’ve adequately prepared for one is very
important. How long will a piece of data be accessible, and by whom? How will the
data be stored and protected over its life? As we mentioned previously, data purging
needs to happen according to the rules set forth. In addition, understanding what
your systems’ backup and retention policies are, is important.
Guidance 4: Define data policies
It’s important to document how data will be managed and shared. Identify the licens‐
ing and sharing agreements that pertain to the data you’re collecting. Are there
restrictions that the organization should adhere to? What are the legal and ethical
restrictions on access to and use of sensitive data, for example? Regulations like
GDPR and CCPA can easily get confusing and can even become contradictory. In this
Data Life Cycle Management
|
91
step, ensure that all the applicable data policies are captured accordingly. This also
helps in case you’re audited.
Guidance 5: Define roles and responsibilities
Chapter 3 defined roles and responsibilities. With those roles in mind, determine
which are the right ones for your organization and what each one means for you.
Which teams will be responsible for metadata management and data discovery? Who
will ensure governance policies are followed all the way? And there are many more
roles that you can define.
A DMP should provide your organization with an easy-to-follow roadmap that will
guide others and explain how data will be treated throughout its life cycle. Think of
this as a living document that evolves with your organization as new datasets are cap‐
tured, and as new laws and regulations are enacted.
If this was a data management plan for a research project, it would have included a
lot more steps and items for consideration. Those plans tend to be more robust
because they guide the entire research project and data end-to-end. We will cover a
lot more concepts later in the chapter, so we chose to select items that were easily
transferable to creating a governance policy and plan for your organization.
Applying Governance over the Data Life Cycle
We’ve gone through fundamental concepts thus far; now let’s bring everything
together and look at how you can apply governance over the data life cycle. Gover‐
nance needs to bring together people, processes, and technology to govern data
throughout its life cycle. In Chapter 2, we outlined a robust set of tools to make gov‐
ernance a reality, and Chapter 3 focused on the people and process side of things. It’s
important to point out that implementing governance is complicated; there’s no easy
way to simply apply everything and consider he job done. Most technologies need to
be stitched together, and as you can imagine, they’re all coming from different ven‐
dors with different implementations. You would need to integrate the best-in-class
suite of products and services to make things work. Another option is to purchase a
fully integrated data platform or governance platform. This is not a trivial task.
Data Governance Framework
Frameworks help you visualize the plan, and there are several frameworks that can
help you think about governance across the data life cycle. Figure 4-2 is one such
framework in which we highlight all the concepts from Chapter 2, overlaid with the
concepts we’ve discussed in this chapter.
92
| Chapter 4: Data Governance over a Data Life Cycle
Figure 4-2. Governance over a data life cycle
This framework oversimplifies things to make them easier to understand; it assumes
things are linear, from left to right, which is usually not the case. When data is inges‐
ted from various sources on the left, this is simply at the point of data creation or cap‐
ture. That data is then processed and stored, and then it is consumed by the different
stakeholders, including data analysts, data engineers, data stewards, and so on.
Data archiving and data destruction are not reflected in this framework because those
take place beyond the point when data is used. As we previously outlined, during
archiving, data is removed from all active production environments. It is no longer
processed, used, or published but is stored in case it is needed again in the future.
Destruction is when data comes to the end of its life and is removed according to
guidelines and procedures that have been set forth.
One discrepancy that you will quickly notice is that metadata management should be
considered from the point of data creation—where enterprises need to discover and
curate the data as it’s ingested (especially for sensitive data)—to when data is stored
and discovered in the applicable storage system. Archiving, even though mentioned
within data management, tends to happen when the data’s usefulness is done and it is
removed from production environments. Though archiving is an important part of
governance, this diagram implies that it is taking place in the middle of the data life
cycle. That said, it’s also possible to have an archiving strategy when data is simply
stored in the applicable storage systems, so we cannot completely rule this out.
Applying Governance over the Data Life Cycle
|
93
We want to reiterate that Figure 4-2 provides a logical representation of the phases a
piece of data goes through, from left to right, and not necessarily the actual step-bystep flow of the data. There’s a lot of back and forth that happens between each phase,
and not all pieces of data go through each one of these phases.
Frameworks are good at providing a holistic view of things. However, they are not the
be-all, end-all. Make sure whichever framework you select works for your organiza‐
tion and your data.
We’ve mentioned it already, but would like to emphasize again the
idea of selecting a framework that work for your organization. This
can include considerations around the kind of data you collect or
work with, as well as what kind of personnel you have dedicated to
your data governance efforts. One thing we challenge you to con‐
sider is how to take what you have and fit enough framework
around it. Take these ideas as laid out (noting that not each step is
required or even necessary) and layer on what you have to work
with currently as a place to start. More pieces (and people, for that
matter) can be added later, but if you focus on at least laying the
groundwork—the foundation—you will be in a much better posi‐
tion if or when you do have more pieces to add to your framework.
Data Governance in Practice
OpenStreetMap (OSM) was created by Steve Coast in the UK in 2004 and was
inspired by the success of Wikipedia. It is open source, which means it is created by
people like you and is free to use under an open license. It was a response to the pro‐
liferation of siloed, proprietary international geographical data sources and dozens of
mapping software products that didn’t talk to each other. OSM has grown signifi‐
cantly to over two million contributors, and what’s amazing is that it works. In fact, it
works well enough to be the trusted source of data for a number of Fortune 500 com‐
panies, including other small and medium-sized businesses. With so many contribu‐
tors, OSM is successful because it was able to establish data standards early in the
process and ensured contributors adhered to them. As you can imagine, a crowd‐
sourced mapping system without a way to standardize contributor data could go
wrong very quickly. Defining governance standards can bring value to your organiza‐
tion and provide trusted data for your users.
And now that you have an understanding of the data life cycle with an overlay of the
different governance tools, let’s delve further into how the different data governance
tools we outlined in Chapters 1 and 2 can be applied and used across this life cycle.
This section also includes best practices, which can help you start to define your
organization’s data standards.
94
| Chapter 4: Data Governance over a Data Life Cycle
Data creation
As previously mentioned, this is the initial phase of the data life cycle, where data is
created or captured. During this phase, an organization can choose to capture both
the metadata, and the lineage of the data. Metadata describes the data, while the line‐
age describes the where of the data and how it will flow and be transformed and used
downstream. Trying to capture these during this initial phase sets you up well for the
later phases.
In addition, processes such as classification and profiling can be employed, especially
if you’re dealing with sensitive data assets. Data should also be encrypted in transit to
offer protection from intrusions and attacks. Cloud service providers such as Google
Cloud offer encryption in transit and at rest by default.
Define the Type of Data
Establish a set of guidelines for categorizing data that takes into
account the sensitivity of the information as well its criticality and
value to the organization. Profiling and classifying data helps
inform which governance policies and procedures apply to the
data.
Data processing
During this phase, data goes through processes such as integration, cleaning, scrub‐
bing, or extract-transform-load (ETL) prior to its use, to get it ready for storage and
eventual analysis. It’s important that the integrity of the data is preserved during this
phase; that is why data quality plays a critical role.
Lineage needs to be captured and tracked here also, to ensure that the end users
understand which processes led to which transformation and where the data origina‐
ted from. We heard this from one user: “It would be nice to have a better understand‐
ing of the lineage of data. When finding where a certain column in a table comes
from, I need to manually dig through the source code of that table and follow that
trail (if I have access). Automate this process.” This is a common pain point felt by
many, and one in which DLM and governance are critical.
Document Data Quality Expectations
Different data consumers may have different data quality require‐
ments, so it’s important to provide a means to document data qual‐
ity expectations while the data is being captured and processed, as
well as techniques and tools for supporting the data’s validation and
monitoring as it goes through the data life cycle. The right pro‐
cesses for data quality management will provide measurable and
trustworthy data for analysis.
Applying Governance over the Data Life Cycle
|
95
Data storage
In this phase, both data and metadata are stored and made ready for analysis. Data
should be encrypted at rest to protect it from intrusions and attacks. In addition, data
needs to be backed up to ensure redundancy.
Automated Data Protection and Recovery
Because data is stored in storage devices in this phase, find solu‐
tions and products that provide automated data protection to
ensure that exposed data cannot be read, including encryption at
rest, encryption in transit, data masking, and permanent deletion.
In addition, implement a robust recovery plan to protect your busi‐
ness when a disaster strikes.
Data usage
In this phase, data is analyzed and consumed for insights and consumed by multiple
internal and external stakeholders and processes in the organization. In addition,
analyzed data is visualized and used to support the organization’s objectives and oper‐
ations; business intelligence tools play a critical role in this phase.
A data catalog is vital to helping users discover data assets using captured metadata.
Privacy, access management, and auditing are paramount at this stage, which ensures
that the right people and systems are accessing and sharing the data for analysis. Fur‐
thermore, there might be regulatory or contractual constraints on how data may
actually be used, and part of the role of data governance is to ensure that these con‐
straints are observed.
Data Access Management
It is important to provide data services that allow data consumers
to access their data with ease. Documenting what and how the data
will be used, and for what purposes, can help you define identities,
groups, and roles and assign access rights to establish a level of
managed access. This ensures that only authorized and authentica‐
ted individuals and systems are able to access data assets according
to defined rules.
Data archiving
In this phase, data is removed from all active production environments. It is no
longer processed, used, or published but is stored in case it is needed again in the
future. Data classification should guide the retention and disposal method of data.
96
| Chapter 4: Data Governance over a Data Life Cycle
Automated Data Protection Plan
Beyond being a way to prevent unauthorized individuals from
accessing data, perimeter security is not and never has been suffi‐
cient for protecting data. The same protections applied in data stor‐
age would apply here as well to ensure that exposed data cannot be
read, including encryption at rest, data masking, and permanent
deletion. In addition, in case of a disaster, or in the event that
archive data is needed again in a production environment, it’s
important to have a well-defined process to revive this data and
make it useful.
Data destruction
Finally, data is destroyed, or rather, it is removed from the enterprise at the end of its
useful life. Before purging any data, it is critical to confirm whether there are any pol‐
icies in place that would require you to retain the data for a certain period of time.
Data classification should guide the retention and disposal method of data.
Create a Compliance Policy
Coming up with the right timeline for this cycle means under‐
standing state and federal regulations, industry standards, and gov‐
ernance policies and staying up to date on any changes. Doing so
helps to ensure that the right steps are taken and that the purge has
been done properly. It also ensures that data doesn’t consume more
resources than necessary at the end of its useful life.
IT stakeholders are urged to revisit the guidelines for destroying
data every 12–18 months to ensure compliance, since rules change
often.
Example of How Data Moves Through a Data Platform
Here’s an example scenario of how data could move through a data platform with the
framework in Figure 4-2.
Scenario
Let’s say that a business wants to ingest data onto a cloud-data platform, like Google
Cloud, AWS, or Azure, and share it with data analysts. This data may include sensi‐
tive elements such as US social security numbers, phone numbers, and email
addresses. Here are the different pieces it might go through:
Applying Governance over the Data Life Cycle
|
97
1. Business configures an ingestion data pipeline using a batch or streaming service:
a. Goal: As they move raw data into the platform, it will need to be scanned,
classified, and tagged before it can be processed, manipulated, and stored.
b. Staged ingestion buckets:
i. Ingest: heavily restricted
ii. Released: processed data
iii. Admin quarantine: needs review
2. Data is then scanned and classified for sensitive information such as PII.
3. Some data may be redacted, obfuscated, or anonymized/de-identified. This pro‐
cess may generate new metadata, such as what keys were used for tokenization.
This metadata would be captured at this stage.
4. Data is tagged with PII tags/labels.
5. Aspects of data quality can be accessed—that is, are there any missing values, are
primary keys in the right format, etc.
6. Start to capture data provenance information for lineage.
7. As data moves between the different services along the life cycle, it is encrypted
in transit.
8. Once ingestion and processing are complete, the data will need to be stored in a
data warehouse and/or a data lake, where it is encrypted at rest. Backup and
recovery processes need to be employed as well, in case of a disaster.
9. While in storage, additional business and technical metadata can be added to the
data and cataloged, and users need to be able to discover and find the data.
10. Audit trails need to be captured throughout this data life cycle and made visible
as needed. Audits allow you to check the effectiveness of controls in order to
quickly mitigate threats and evaluate overall security health.
11. Throughout this process, it is important to ensure that the right people and serv‐
ices have access and permissions to the right data across the data platform using a
robust identity and access management (IAM) solution.
12. You need to be able to run analytics and visualize the results for use. In addition
to access management, additional privacy, de-identification, and anonymization
tools may be employed.
13. Once this data is no longer needed in a production environment, it is archived
for a determined period of time to maintain compliance.
14. At the end of its useful life, it is completely removed from the data platform and
destroyed.
98
|
Chapter 4: Data Governance over a Data Life Cycle
Case Study: Nike and the 4% Improving Running Shoes
In 2018, Nike launched a new running shoe, and the related advertising campaign
claimed that the “Nike Zoom Vaporfly 4%” will make you run 4% faster. While a 4%
speed increase does not sound like a lot, over a marathon run that averages to 4–5
hours, this can potentially lead to 10–12 minutes’ improvement in the time it takes to
finish the race.
The claims were significant and were met with skepticism. A Nike-sponsored study
did not help because it was based, by necessity, on a small dataset. Many athletes
would pay for an expensive shoe to improve their result by that margin, if the claim
were true, but the source of this information (the vendor) sowed doubt. Running an
independent scientific experiment to conclude whether or not the claim was true
would have been challenging, as this would have required significant investment, and
getting the runner(s) to use different kinds of shoes over the same courses in the same
conditions to truly eliminate all possible variables and challenges.
Happily, many athletes use a popular record-keeping app called Strava. Strava makes
the data publicly available, and many athletes also record the shoes they use when
running (Figure 4-3). This created a natural experiment in which you could look over
existing data and, with enough data, could potentially tease out patterns.
Figure 4-3. Strava data. Does the Nike Vaporfly make you run 4% faster?
The New York Times investigated, collecting half a million real-life performance
records from Strava.Keven Quealy and Josh Katz, “Nike Says Its $250 Running Shoes
Will Make You Run Much Faster. What If That’s Actually True?” New York Times, July
18, 2018. The next step was to determine whether or not the data was useful. While
the ideal way would have been to measure runs of identical runners on the same
course with different shoes, this was not possible at this scale. However, the large
Applying Governance over the Data Life Cycle
|
99
amount of data did enable the finding of natural experiments.2 This was not a smallscale lab experiment, but an actual account of (admittedly amateur, for the most part)
weekend runners reporting and sharing their race results.
The New York Times was able to compile 280,000 marathon and 215,000 half mara‐
thon results, and then compared running conditions (same races) and concurrent
results from the same athletes (different shoes, different races or dates). These com‐
parisons ensured that conditions similar to the ideal experiment were met, and by
including knowledge about the races themselves (weather, course layout and diffi‐
culty), the data was curated to keep the quality records in while eliminating outliers
(less-frequented races, extreme conditions).
The New York Times was able to conclude that the Nike shoes were very often part of
a more successful run for many runners. The paper pointed out that these results
were not gathered through a lab experiment in controlled conditions, but their results
were consistent with the Nike-funded study.
This effort would not have been possible without the availability of an open dataset,
contributed to freely by runners globally, and preserved and made accessible to
researchers. This example of a dataset in which data is made available under a con‐
trolled environment (Strava does protect runners’ personal data and allows runners
full control over how their data is shared, including the ability to opt out and delete
their own data) is an excellent example of proper information cycle and data
governance.
Operationalizing Data Governance
It’s one thing to have a plan, but it’s something else to ensure that plan works for your
organization. NASA learned things the hard way. In September 1999, after almost 10
months of travel to Mars, the $125-million Mars Climate Orbiter lost communication
and then burned and broke into pieces a mere 37 miles away from the planet’s sur‐
face. The analysis found out that, while NASA had used the metric system, one of its
partners had used the International System of Units (SI). This inconsistency was not
discovered until it was time to land the orbiter, leading to a complete loss of the satel‐
lite. This of course was crushing to the team. After this incident, proper checks and
balances were implemented to ensure that something similar did not happen again.
Bringing things together so that issues such as the one NASA experienced are caught
early and rectified before a disaster happens starts with the creatiion of a data gover‐
2 A natural experiment is a situation in which you can identify experimental and control groups determined by
factors outside the control of the investigators. In our example here, the runners naturally fell into groups
defined by the shoes they wore, rather than being assigned shoes externally. The groups of runners were large
enough to qualify for good “experimental” and “control” groups with controlled number of external factors.
100
|
Chapter 4: Data Governance over a Data Life Cycle
nance policy. A data governance policy is a living, breathing document that provides
a set of rules, policies, and guidance for safeguarding an organization’s data assets.
What Is a Data Governance Policy?
A data governance policy is a documented set of guidelines for ensuring that an
organization’s data and information assets are managed consistently and used prop‐
erly. A data governance policy is essential in order to implement governance. The guide‐
lines will include individual policies for data quality, access, security, privacy, and
usage, which are paramount for managing data across its life cycle. In addition, data
governance policies center on establishing roles and responsibilities for data that
include access, disposal, storage, backup, and protection, which should all be familiar
concepts. This document helps to bring everything together toward a common goal.
The data governance policy is usually created by a data governance committee or data
governance council, which is made up of business executives and other data owners.
This policy document defines a clear data governance structure for the executive
team, managers, and line workers to follow in their daily operations.
To get started operationalizing governance, a data governance charter template could
be useful. Figure 4-4 shows an example template that could help you socialize your
ideas across the organization and get the conversation started. Information in this
template will funnel directly into your data governance policy.
Use the data governance charter template to kick off the conversation and get your
team assembled. Once it has bought into your vision, mission, and goals, that is the
team that will help you create and define your governance policy.
Operationalizing Data Governance
|
101
Figure 4-4. Data governance charter template
Importance of a Data Governance Policy
When you have a business idea and are going to friends to socialize the idea and pos‐
sibly get them to buy into it, you will quickly run into someone who asks for a busi‐
ness plan. “Do you have a business plan you can share so I can read more about this
idea and what your plans are?” A data governance policy allows you to have all the
important elements of operationalizing governance documented according to your
organization’s needs and objectives. It also allows consistency within the organization
over a long period of time. It is the document that everyone will refer to when ques‐
tions and issues arise. It should be reviewed regularly and updated when things in the
organization change. You can consider it your business plan—or to another extreme,
it can also be your governance bible.
When a data governance policy is well drafted, it will ensure:
• Consistent, efficient, and effective management of the data assets throughout the
organization and data life cycle and over time.
• The appropriate level of protection of the organization’s data assets based on their
value and risk as determined by the data governance committee.
• The appropriate protection and security levels for different categories of data as
established by the governance committee.
102
|
Chapter 4: Data Governance over a Data Life Cycle
Developing a Data Governance Policy
A data governance policy is usually authored by the data governance committee or
appointed data governance council. This committee will establish comprehensive
policies for the data program that outline how data will be collected, stored, used, and
protected. The committee will identify risks and regulatory requirements and look
into how they will impact or disrupt the business.
Once all the risks and assessments have been identified, the data governance commit‐
tee will then draft policy guidelines and procedures that will ensure the organization
has the data program that was envisioned. When a policy is well written, it helps cap‐
ture the strategic vision of the data program. The vision for the governance program
could be to drive digital transformation for the organization, or possibly to get
insights to drive new revenue or even to use data to provide new products or services.
Whichever is the case for your organization, the policies drafted should all coalesce
toward the articulated vision and mission as outlined in the data governance charter
template.
Part of the process of developing a data governance policy is establishing the expecta‐
tions, wants, and needs of key stakeholders through interviews, meetings, and infor‐
mal conversations. This will help you get valuable input, but it’s also an opportunity
to secure additional buy-in for the program.
Data Governance Policy Structure
A well-crafted policy should be unique to your organization’s vision, mission, and
goals. Don’t get hung up on every single piece of information on this template, how‐
ever; use it more like a guide to help you think things through. With that in mind,
your governance policy should address:
Vision and mission for the program
If you used a data governance charter template as outlined in Figure 4-4 to get
buy-in from other stakeholders, that means you already have this information
readily available. As mentioned before, the vision for the governance program
could be to drive digital transformation for the organization, or to get insights to
drive new revenue, or even to use data to provide new products or services.
Policy purpose
Capture goals for your organization’s data governance program, as well as metrics
for determining success. The mission and vision of the program should drive the
goals and success metrics.
Policy scope
Document the data assets covered by this governance policy. In addition, inven‐
tory the data sources and determine data classifications based on whether data is
Operationalizing Data Governance
|
103
sensitive, confidential, or publicly available, along with the levels of security and
protection required at the different levels.
Definitions and terms
The data governance policy is usually viewed by stakeholders across the organi‐
zation who might not be familiar with certain terms. Use this section to docu‐
ment terms and definitions to ensure everyone is on the same page.
Policy principles
Define rules and standards for the governance program you’re looking to set up
along with the procedures and programs to enforce them. The rules could cover
data access (who has access to what data), data usage (how the data will be used
and details around what’s acceptable), data integration (what transformations the
data will undergo), and data integrity (expectations around data quality).
Develop best practices to protect data and to ensure regulations and compliance
are effectively documented.
Program structure
Define roles and responsibilities (R&Rs), which are positions within the organi‐
zation that will oversee elements of the governance program. A RACI chart could
help you map out who is responsible, who is accountable, who needs to be con‐
sulted, and who should be kept informed about changes. Information on gover‐
nance R&Rs can be found in Chapter 3 of the book.
Policy review
Determine when the policy will be reviewed and updated and how adherence to
the policy will be monitored, measured, and remedied.
Further assistance
Document the right people to address questions from the team and other stake‐
holders.
It’s not enough to document a data governance policy as outlined in Figure 4-5, com‐
municating it to all stakeholders is equally important. This could happen through a
combination of group meetings and training, one-on-one conversations, recorded
training videos, and written communication.
104
|
Chapter 4: Data Governance over a Data Life Cycle
Figure 4-5. Example data governance policy template
In addition, review performance regularly with your data governance team to ensure
that you’re still on the right track. This also means regularly reviewing your data gov‐
ernance policy to make sure it still reflects the current needs of the organization and
program.
Roles and Responsibilities
When operationalizing governance over a data life cycle, you will interact with many
stakeholders within the organization, and you will need to bring them together to
work on this common goal. While it might be tempting to definitively say which roles
do what at which part of the data life cycle, as outlined in Chapter 3, many data gov‐
ernance frameworks revolve around a complex interplay of roles and responsibilities.
The reality is that most companies rarely are able to exactly or fully staff governance
roles due to lack of employee skill set or, more commonly, a simple lack of headcount.
For this reason, employees working in the information and data space of their com‐
pany often wear different user “hats.”
We will not go into detail about roles and responsibilities in this chapter, because
they’re well outlined in Chapter 3. You still need to define what these look like within
your organization and how they will interplay with each other to make governance a
reality for you. This will typically be outlined in a RACI matrix describing who is
“responsible, accountable, to be consulted, and to be informed” within a certain
enforcement, process, policy, or standard.
Operationalizing Data Governance
|
105
Step-by-Step Guidance
By this section of the book, you should know that data governance goes beyond the
selection and implementation of products and tools. The success of a data governance
program depends on the combination of people, processes, and tools all working
together to make governance a reality. This section will feel very familiar, because it
gathers all the elements discussed in the previous section on data governance policy
and puts them in a step-by-step process to show you how to get started. It further
double clicks into the concepts as well.
Build the business case
As previously mentioned, data governance takes time and is expensive. If done
correctly, it can be automated as part of the application design done at the source
with a focus on business value. That said, data governance initiatives will often
vary in scope and objectives. Depending on where the initiative is originating
from, you need to be able to build a business case that will identify critical busi‐
ness drivers and justify the effort and investment of data governance. It should
identify the pain points, outline perceived data risks, and indicate how gover‐
nance helps the organization mitigate those risks and enable better business out‐
comes. It’s OK to start small, strive for quick wins, and build up ambitions over
time. Set clear, measurable, and specific goals. You cannot control what you can‐
not measure; therefore you need to outline success metrics. The data governance
charter template in Figure 4-4 is perfect for helping you get started.
Document guiding principles
Develop and document core principles associated with governance and, of
course, associated with the project you’re looking to get off the ground. A core
principle of your governance strategy could be to make consistent and confident
business decisions based on trustworthy data aligned with all the various pur‐
poses for the use of the data assets. Another core principle could be to meet regu‐
latory requirements and avoid fines or even to optimize staff effectiveness by
providing data assets that meet the desired data quality thresholds. Define princi‐
ples that are core to your business and project. If you’re still new to this area,
there are a lot of resources available. If you are looking online, there are several
vendor-agnostic, not-for-profit associations, such as the Data Governance Insti‐
tute (DGI), the Data Management Association (DAMA), the Data Governance
Professionals Organization (DGPO), and the Enterprise Data Management
Council, all of which provide great resources for business, IT, and data professio‐
nals dedicated to advancing the discipline of data governance. In addition, iden‐
tify whether there are any local data governance meetup groups or conferences
that you can possibly attend, such as the Data Governance and Information
Quality Conference, DAMA International Events, or a Financial Information
Summit.
106
|
Chapter 4: Data Governance over a Data Life Cycle
Get management buy-in
It should be no surprise that without management buy-in, your governance ini‐
tiative can easily be dead from the get-go. Management controls the big decisions
and funding that you need. Outlining important KPIs, and how your plan helps
to move them, will get management to be all ears. Engage data governance cham‐
pions and get buy-in from the key senior stakeholders. Present your business
case and guiding principles to C-level management for approval. You need allies
on your side to help make the case. And once the project has gotten off the
ground, communicate frequently.
Develop an operating model
Once you have management approval, it’s time to get to work. How do you inte‐
grate this governance plan into the way of doing business in your enterprise? We
introduced you to the data governance policy, which can come in very handy
during this process. During this stage, define the data governance roles and
responsibilities, and then describe the processes and procedures for the data gov‐
ernance council and data stewardship teams who will define processes for defin‐
ing and implementing policies as well as for reviewing and remediating identified
data issues. Leverage the content from the data management policy plan to help
you define your operating model. Data governance is a team sport, with delivera‐
bles from all parts of the business.
Develop a framework for accountability
As with any project you’re looking to bring to market, establishing a framework
for assigning custodianship and responsibility for critical data domains is para‐
mount. Define ownership. Make sure there is visibility to the “data owners”
across the data landscape. Provide a methodology to ensure that everyone is
accountable for contributing to data usability. Refer back to your data manage‐
ment policy, as it probably started to capture some of these dependencies.
Develop taxonomies and ontologies
This is where a lot of the education you’ve collected thus far comes in handy.
Working closely with governance associations, leaning in on your peers, and sim‐
ply learning about things online will help you with this step. There may be a
number of governance directives associated with data classification, organization,
and, in the case of sensitive information, data protection. To enable your data
consumers to comply with those directives, there must be a clear definition of the
categories (for organizational structure) and classifications (for assessing data
sensitivity). These should be captured in your data governance policy.
Assemble the right technology stack
Once you’ve assigned data governance roles to your staff and defined and
approved your processes and procedures, you should then assemble a suite of
tools that facilitates implementation and ongoing validation of compliance with
Operationalizing Data Governance
|
107
data policies and accurate compliance reporting. Map infrastructure, architec‐
ture, and tools. Your data governance framework must be a sensible part of your
enterprise architecture, the IT landscape, and the tools needed. We talked about
technology in previous sections, so we won’t go into detail about it here. Finding
tools and technology that work for you and satisfy the organizational objectives
you laid out is what’s important.
Establish education and training
As highlighted earlier, for data governance to work, it needs buy-in across the
organization. You need to ensure that your organization is keeping up and is still
buying into the project you presented. It’s therefore important to raise awareness
of the value of data governance by developing educational materials highlighting
data governance practices, procedures, and the use of supporting technology.
Plan for regular training sessions to reinforce good data governance practices.
Wherever possible, use business terms, and translate the academic parts of the
data governance discipline into meaningful content in the business context.
Considerations for Governance Across a Data Life Cycle
Data governance has been around since there was data to govern, but it was mostly
viewed as an IT function. Implementing data governance across the data life cycle is
no walk in the park. Here are some considerations you will need to think about as
you implement governance in your organization. These should not be surprising to
you, because you will quickly notice that they touch on a lot of aspects we introduced
in Chapters 1 and 2, as well as in this chapter.
Deployment time
Crafting and setting up governance processes across the data life cycle takes a lot of
time, effort, and resources. In this chapter, we have introduced a lot of concepts,
ideas, and ways to think about operationalizing governance across the data life cycle,
and you can see it gets overwhelming very quickly. There’s not a one-size-fits-all solu‐
tion; you need to identify what is unique about your business and then forge a plan
that works for you. Automation can reduce the deployment time compared with
hand-coded governance processes. In addition, artificial intelligence is seen as a way
to get arms around data governance in the future, especially for things like autodis‐
covery of sensitive data and metadata management. That means that as you look for
solutions in the market, you will need to find out how much automation and integra‐
tion is built into it, how well it works for your environment and situation, and
whether that is the most difficult part of the workflow that could use automation. In a
hybrid or even a multi-cloud world, this becomes even more complex and further
increases the deployment time.
108
|
Chapter 4: Data Governance over a Data Life Cycle
Complexity and cost
Complexity comes in many forms. In Chapter 1, we talked about how much the data
landscape is and about just how quickly data was being produced in the world.
Another complexity is a lack of defined industry standards for things like metadata.
We touched on this in Chapter 2. In most cases, metadata does not obey the same
policies and controls as the underlying data itself, and a lack of standardized meta‐
data specifications means that different products and processes will have different
ways of presenting this information. Still another complexity is the sheer amount of
tools, processes, and infrastructure needed to make governance a reality. In order to
deliver comprehensive governance, organizations must either integrate best-of-breed
solutions, which are often complex and very expensive (with high license and mainte‐
nance costs), or buy turnkey, integrated solutions, which are expensive and fewer in
the market. With this in mind, cloud service providers (CSPs) are building data plat‐
forms with all these governance capabilities built in, thus creating a one-stop shop
and simplifying the process for customers. As an organization, research and compare
the different data platforms provided by CSPs and see which one works for you. Some
businesses choose to leave some of their data on-premises; however, for the data that
can move to the cloud, these CSPs are now building robust tools and processes to
help customers govern their data end-to-end on the platform. In addition, companies
such as Informatica, Alation, and Collibra offer governance-specific platforms and
products that can be implemented in your organization.
Changing regulation environment
In previous chapters, we’ve clearly outlined the implications of a constantly changing
regulatory environment with the introduction of GDPR and CCPA. We will not go
into the same detail here; however, regulations define a lot of what must be done and
implemented to ensure governance. They will outline how certain types of data need
to be handled and which types of controls need to be in place, and they sometimes
will even go as far as outlining what the repercussions are when these things are not
complied with. Complying with regulations is absolutely something your organiza‐
tion needs to think about as you implement data governance over the data life cycle.
Operationalizing Data Governance
|
109
In our discussions with many different companies, we’ve heard of
two very different philosophies when it comes to considering
changes to the regulatory environment. One strategy is to assume
that, in the future, the most restrictive regulations that are present
now will cascade and be required everywhere (like CCPA being
required across the entire US and not just in California), and that
ensuring compliance now, even though not required, is a top prior‐
ity. Conversely, we’ve also heard the strategy of complying only
with what’s required right now and dealing with regulations only if
they become required. We strongly suggest you take the former
approach, because a proper and well-thought-out governance pro‐
gram not only ensures compliance with ever-changing regulations;
it also enables many of the other benefits we’ve outlined thus far,
such as better findability, better security, and more accurate analyt‐
ics from higher-quality data.
Location of data
In order to fully implement governance over a data life cycle, understanding which
data is on-premises versus in the cloud is very important. Furthermore, understand‐
ing how data will interact with other data along the life cycle does create complexity.
In the current paradigm, most organizational data lives both on-premises and in the
cloud, and having systems and tools that allow for hybrid and even multicloud sce‐
narios is paramount. In Chapter 1, we talked about why governance is easier in the
public cloud—it’sprimarily because the public cloud has several features that make
data governance easier to implement, monitor, and update. In many cases, these fea‐
tures are unavailable or cost-prohibitive in on-premises systems. Data should be pro‐
tected no matter where it is located, so a viable data life cycle management plan will
incorporate governance for all data at all times.
Organizational culture
As you know, culture is one of those intangible things in an organization that plays an
important role in how the organization functions. In Chapter 3, we touched on how
an organization can create a culture of privacy and security, which allows employees
to understand how data should be managed and treated so that they are good stew‐
ards of proper data handling and usage. In this section, we’re referring to organiza‐
tional culture, which often dictates what people do and how they behave. Your
organization might be free, allowing folks to easily raise questions and concerns, and
in such an environment, when something goes wrong, people are more likely to speak
up. In organizations in which people are reprimanded for every little thing, they will
be more afraid to speak up and report when things are not working or even when
things go wrong. In these environments, governance is a little difficult to implement,
because without transparency and proper reporting, mistakes are usually not discov‐
ered until much later. In the NASA example we provided earlier in this chapter, there
110
|
Chapter 4: Data Governance over a Data Life Cycle
were a couple of people within the organization who noticed the discrepancy in the
data and even reported it. Their reports were ignored by management, and we all
know what happened. Things did not end well for NASA. Remember, instituting gov‐
ernance in an organization is often met with resistance, especially if the organization
is accustomed to decentralized operations. Creating an environment in which func‐
tions are centralized across the data life cycle simply means that these areas have to
adhere to processes that they might not have been used to in the past but that are for
the larger good of the organization.
Summary
Data life cycle management is paramount to implementing governance and ensures
that useful data is clean, accurate, and readily available to users. In addition, it
ensures that your organization remains compliant at all times.
In this chapter, we introduced you to data life cycle management, and how to apply
governance over the data life cycle. We then looked into operationalizing governance
and how the role of a data governance policy is to ensure that an organization’s data
and information assets are managed consistently and used properly. Finally, we pro‐
vided step-by-step guidance for implementing governance and finished with the con‐
siderations for governance across the data life cycle, including deployment time,
complexity and cost, and organizational culture.
Summary
|
111
CHAPTER 5
Improving Data Quality
When most people hear the words data quality, they think about data that is correct
and factual. In data analytics and data governance, data quality has a more nuanced
set of qualifiers. Being correct is not enough, if all of the details are not available (e.g.,
fields in a transaction). Data quality is also measured in the context of a use case, as
we will explain. Let’s begin by exploring the characteristics of data quality.
What Is Data Quality?
Put simply, data quality is the ranking of certain data according to accuracy, com‐
pleteness (all columns have values), and timeliness. When you are working with large
amounts of data, the data is usually acquired and processed in an automated way.
When thinking about data quality, it is good to discuss:
Accuracy
Whether the data captured was actually correct. For example, an error in data
entry causing multiple zeros to be entered ahead of a decimal point, is an accu‐
racy issue. Duplicate data is also an example of inaccurate data.
Completeness
Whether all records captured were complete—i.e., there are no columns with
missing information. If you are managing customer records, for example, make
sure you capture or otherwise reconcile a complete customer details record (e.g.,
name/address/phone number). Missing fields will cause issues if you are looking
for customer records in a specific zip code, for example.
Timeliness
Transactional data is affected by timeliness. The order of events in buying and
selling shares, for example, can have an impact on the buyer’s available credit.
Timeliness also should take into account the fact that some data can get stale.
113
In addition, the data quality can be affected by outlier values. If you are looking at
retail transactions, for example, very large purchase sums are likely indications of
data-entry issues (e.g., forgetting a decimal point) and not indicators that revenues
went up by two orders of magnitude. This will be an accuracy issue.
Make sure to take all possible values into account. In the above retail example, nega‐
tive values are likely indications of returns and not “purchasing a product for nega‐
tive $” and should be accounted for differently (e.g., a possible impact will be the
average transaction size—with the purchase and the return each accounting for a sin‐
gle purchase).
Finally, there is the trustworthiness of the source of the data. Not all data sources are
equal—there is a difference, for example, between a series of temperature values col‐
lected over time from a connected thermometer and a series of human reads of a
mercury thermometer collected over time in handwriting. The machine will likely
control variables such as the time the sample was taken and will sync to a global
atomic clock. The human recording in a notebook will possibly add variance to the
time the sample was taken, may smudge the text, or may have hard-to-read handwrit‐
ing. It is dangerous to take data from both sources and treat them as the same.
Why Is Data Quality Important?
For many organizations, data directly leads to decision making: a credit score com‐
piled from transaction data can lead to a decision by a banker to approve a mortgage.
A company share price is instantly computed from the amount offered by multiple
buyers and sellers. These kinds of decisions are very often regulated—clear evidence
must be collected on credit-related decisions, for example. It is important to the cus‐
tomer and the lender that mortgage decisions are made according to high-quality
data. Lack of data quality is the source of lack of trust and of biased, unethical auto‐
mated decisions being made. A train schedule you cannot trust (based on erroneous
or untimely station visits and past performance) can lead to you making decisions
about your commute that can potentially result in your always taking your own car,
thus negating the very reason for the existence of the mass transit train.
When collecting data from multiple sources or domains, data accuracy and context
become a challenge: not only is it possible that the central repository does not have
the same understanding of the data as the data sources (e.g., how outliers are defined
and how partial data is treated), but it is also possible that data sources do not agree
with each other on the meaning of certain values (e.g., how to handle negative values,
or how to fill in for missing values). Reconciliation of data meaning can be done by
making sure that when a new data source is added, the accuracy, completeness, and
timeliness of the data in the data source is examined—sometimes manually—and
either described in a way that the data analysts using the resource can use or directly
normalized according to the rules in the central repository.
114
|
Chapter 5: Improving Data Quality
When an error or unexpected data is introduced into the system, there are usually no
human curators who will detect and react to it. In a data processing pipeline, each
step can introduce and amplify errors in the next steps, until presented to the busi‐
ness user:
• In the data-gathering endpoints, gathering data from low-quality sources which
are potentially irrelevant to the business task can, if not eliminated early, cause
issues. Consider, for example, mobile ad impressions, or usage details, where
some of the data is gathered for an engineering lab source and does not represent
real users (and can potentially be very large in volume).
• In the landing zone step, upon normalization/aggregation, the wrong data can be
aggregated into a summation and sway the result.
• In the data analytics workload, joining tables of different qualities can introduce
unexpected outcomes.
All of the above can be presented in a way such that the business user is kept unaware
of decisions/operations happening earlier in the data acquisition chain (see
Figure 5-1) and is presented with the wrong data.
Figure 5-1. Simple data acquisition chain
Any step in the (very simple) chain shown in Figure 5-1 can result in erroneous data,
which will eventually drive the wrong business decisions. Let’s look at a couple of
examples.
In the early days of internet mapping, the mapping provider of the time, a company
called MaxMind, was the sole provider of IP addresses to location services. This com‐
pany made an arguably reasonable decision to have a “default” location right at the
center of the United States, in northern Kansas near the Nebraska border. From the
time of making this decision until very recently, whenever the service could not find a
map location for an IP address, it would provide this default location. The problem
with this decision became evident when systems and persons downstream did not
realize the meaning of this default, and when illegal activity was detected from
Why Is Data Quality Important?
|
115
“default” (aka unknown) IP addresses—law enforcement would show up at this loca‐
tion in the central US and serve warrants to people who actually live there.1
A more current data quality challenge example is within California’s collection of
COVID-19 cases. California moved from rapid per-county reports to state data
(which caused an inconsistency). Later, California moved from “people tested” to
“specimens tested” (some people are likely tested more than once, causing another
data quality issue), and then in August 2020 the data quality became an even more
serious issue
after a series of errors—among them, a failed link to one of the country’s largest testing
labs—had led the state to underreport coronavirus cases starting in late July. Counties
were forced to comb spreadsheets in search of reliable data, while the state worked to
understand the scope of the problem and fix it. A total of 296,000 records were
affected.2
It took nearly a month to recover from this data-quality issue.
A study performed by Experian and published by MIT Sloan researchers estimates
that the cost of bad data (and by that they mean bad, or unmanaged data quality) is
around 15–20% of revenue for most companies.3 The study sampled one hundred
“records” (or units of work) from enterprise divisions and then manually calculated
the error range of these records. The result was an astonishing error rate of 50%. The
Data Warehousing Institute (TDWI) estimated in 2002 that poor data quality costs
businesses in the US over $700 billion annually, and since then this figure has grown
dramatically. Currently, IBM estimates that the yearly cost of “bad data” is $3.1
trillion.4
Data Quality in Big Data Analytics
Data warehouses—databases that are used to performing data analytics in petabyte
scale—are vulnerable to data quality issues. Typically, to get data in a big data ware‐
house, you will extract, clean, transform, and integrate data from multiple operational
databases to create a comprehensive database. A set of processes termed extracttransform-load (ETL) is used to facilitate construction of data warehouses. The data
in the warehouses, though rarely updated, is refreshed periodically and is intended
for a read-only mode of operation. The desired use of the data also has an impact on
1 Kashmir Hill, “How an Internet Mapping Glitch Turned a Random Kansas Farm into a Digital Hell”, Splinter,
April 10, 2016.
2 Fiona Kelliher and Nico Savidge, “With Data Backlog Cleared, California Coronavirus Cases Officially
Decreasing, Newsom Says”, Mercury News, August 14, 2020.
3 Thomas Redman, “Seizing Opportunity in Data Quality”, MIT Sloan Management Review, November 27,
2017.
4 IBM, “The Four V’s of Big Data”, Big Data & Analytics Hub (blog).
116
|
Chapter 5: Improving Data Quality
the kinds of analysis and desired uses of the data. A data warehouse built to support
decision making in retail, where transactions are collected hourly and rounded to the
nearest five-cent value, has different quality needs than a data warehouse built to sup‐
port stock trades, where the transaction time needs to be accurate to the
microsecond, and values range from microtransactions of below one cent to transac‐
tions spanning millions of dollars.
Compared to operational database environments, data warehouses pose additional
challenges to data quality. Since data warehouses integrate data from multiple sour‐
ces, quality issues related to data acquisition, cleaning, transformation, linking, and
integration become critical.
We touched on this earlier, but it’s important to keep in mind that
proper data quality management not only assists in the ability to
run big data analytics but also helps to save on cost and prevent the
loss of productivity. The harder it is for analysts to find and use
high-quality data, coupled with extra engineering time spent hunt‐
ing down and solving data issues, the more cost you will incur, and
your analytics output will suffer. The downstream effects and cost
implications of poorly managed data quality should not be dis‐
counted.
Data Quality in AI/ML Models
Of specific note is data quality within AI/ML models. To broadly generalize, machine
learning models work by extrapolating from existing data using a model to predict
future data (e.g., transaction volume). If the input data has errors within it, the
machine learning model will likely amplify these errors. If the machine learning
model is used to make predictions, and those predictions are then further input into
the model (once acted upon), the predictions become the reality, and the machine
learning model becomes compromised because it will generate, by force of a positive
feedback loop, more and more errors.
The data available for building machine learning models is usually divided into three
nonoverlapping datasets: training, validation, and test. The machine learning model is
developed using the training dataset. Next, the validation dataset is used to adjust the
model parameters so that overfitting is avoided. Last, the test dataset is used to evalu‐
ate the model performance. Errors in one or more datasets can lead to a badly trained
machine learning model, which will yield bad output. Note that there is a fine line
between cleaning all the errors out manually (which can be cost prohibitive over a
large amount of records) and allowing some level of error in a robust model.
Why Is Data Quality Important?
|
117
Quality Beats Quantity, Even in AI
It is a truth universally acknowledged that an AI product manager in possession of a
good idea must be in want of copious amounts of data. The resurgence of AI, starting
in around 2014, has been driven by the ability to train ML models on ever-larger
datasets. The explosion of smartphones, the rise of ecommerce, and the prevalence of
connected devices are some of the trends that have fostered an explosion in the
amount of data that a business has available to it when designing new services or opti‐
mizing existing ones. With larger datasets comes the ability to use larger and more
sophisticated AI models. For example, a typical image classification model now has
hundreds of layers, whereas an AI model from the 1990s had only one layer. Such
models are practical because of the availability of custom ML hardware like GPUs and
TPUs, and the ability to distribute the work across many machines in the public
cloud. Thus, any AI product manager with a good idea will be on the hunt to use as
much data as possible. The accuracy of the AI model and its ability to represent the
real world depends on it being trained with the widest, most representative data
possible.
In our effort to gather more data, however, we should be careful to make sure that the
collected data is of good quality. A small amount of high-quality data yields much
better results than does a large amount of low-quality or outright wrong data. A fasci‐
nating example of this second truth (not as universally acknowledged) comes from an
effort to reduce the overhead of testing water meters in northern Italy.
The goal was to identify malfunctioning mechanical water meters based on their
readings alone. Going out to a site to test a water meter (see Figure 5-2) is pretty
expensive; if the team could use AI to identify potential malfunctions from the water
readings alone, it would be a huge cost savings. For example, if a water meter ran
backwards, or if the amount of water read as being consumed was wholly unreasona‐
ble (it was more, perhaps, than the amount of water supplied to the entire neighbor‐
hood), we could be sure the meter was faulty. Of course, this also depended on the
historical water usage at any meter—a water meter attached to a one-bath house with
a small garden would tend to consume a certain amount of water that varies season‐
ally. A large deviation from this amount would be suspicious.
118
|
Chapter 5: Improving Data Quality
Figure 5-2. A mechanical water meter of the sort used in Italy, meant to be read by a
human looking at the dial. Photo courtesy of Andrevruas, Creative Commons License
3.0
The team started out with 15 million readings from 1 million mechanical water
meters. This was enough data to train a recurrent neural network (RNN), the fanciest
time-series prediction method (with less data, they’d have to settle for something like
ARIMA [an autoregressive integrated moving average model]), and so the team got to
work. The data was already numeric, so there was no need for any data preprocessing
—the team could just feed it into the RNN and do lots of hyperparameter tuning. The
idea was that any large differences between what the RNN predicted and what the
meter actually read would be attributable to faulty water meters.
How did that go? The team notes:
Our initial attempts to train a recurrent neural network, without a specific attention
to the quality, and to the limitations, of those data used for training, led to unexpec‐
ted and negative prediction outcomes.5
5 Marco Roccetti et al., “Is Bigger Always Better? A Controversial Journey to the Center of Machine Learning
Design, with Uses and Misuses of Big Data for Predicting Water Meter Failures”, Journal of Big Data 6, no. 70
(2019).
Why Is Data Quality Important?
|
119
They went back to the drawing board and looked more closely at those 15 million
readings that they had. It turned out that there were two problems with the data: erro‐
neous data and made-up data.
First, some of the data was simply wrong. How could a water meter reading have been
wrong? Wouldn’t a customer complain? It turns that the process of a water meter
reading going to the customer consists of three steps:
1. The mechanical water meter being read by the technician onsite
2. The reading being entered from the technician’s logbook into the company ERP
system
3. The ERP system calculating the bill
In the case of errors in either data entry, the customer might call to complain, but
only the bill was corrected. The water meter readings might still remain wrong! Thus,
the AI model was getting trained on wrong data and being told the meters were not
faulty. Such wrong data turned out to be about 1% of all observations, which is about
the proportion of actually faulty water meters. So, if the meter did return such a bad
value, the model was simply tossing a coin—when the historical data did have these
wild swings, including negative values, half the time it was because the observation
was wrong, and the other half of the time it was because the meter was faulty. The
RNN was being trained on noise.
Second, some of the data was simply made up. It costs so much to send a technician
out to a site that sometimes past readings were simply extrapolated and a reasonable
charge was made to the customer. This was then made up the next time by a true
reading. For example, an actual measurement might be made in January, skipped in
March, and caught up in May. The value for March wasn’t real. Thus the AI model
was being trained on data that wasn’t real—31% of the “measurements” were actually
interpolated values. Further, the readjustments added a lot of noise to the dataset.
After correcting all the billing-related errors and the interpolated values, faulty water
meters were detected by the RNN with an accuracy of 85% (a simpler linear regres‐
sion model would have given them 79%). In order to get there, though, they had to
throw away a large percentage of the original data. Quality, in other words, trumped
quantity.
A good data governance regime would have been careful to propagate billing correc‐
tions back to the original source data and to classify measurements and interpolated
values differently. A good data governance regime would have enforced dataset qual‐
ity from the get-go.
120
|
Chapter 5: Improving Data Quality
Why Is Data Quality a Part of a Data Governance Program?
To summarize, data quality is absolutely essential when planning a data program.
Organizations very often overestimate the quality of the data they have and underes‐
timate the impact of bad data quality. The same program that governs data life cycle,
controls, and usage should be leveraged to govern the quality of data (and plan for
the impact and response to bad data quality incidents).
Techniques for Data Quality
Having discussed the importance of data quality, let’s review a few strategies for clean‐
ing up data, assessing quality, and improving data quality. As a general rule, the ear‐
lier in the pipeline that data can be prepared, sanitized, disambiguated, and cleaned,
the better. It is important to note, in parallel, that data processing pipelines are not the
same for different business purposes/different teams; thus it may be hard to clean up
the data upstream, and that task may need to move downstream for the individual
teams. There is a balance here: if further down the data analytics pipeline aggregates
are performed, causing a coarsening of the data, the crucial meaning needed for
cleanup may be lost as discrete values are aggregated. We will highlight three key
techniques for data quality: prioritization, annotation, and profiling.
The Importance of Matching Business Case to the Data Use
Participants in a virtual symposium for the Society of Vertebrate Paleontology in
October 2020 were baffled by the fact that the transcription, captions, and chat mes‐
sages in digital Q&A sessions were oddly garbled. After some investigation, a com‐
mon denominator was determined: the words “bone,” “knob,” and “jerk”—all very
relevant and appropriate when discussing fossils—were banned and made invisible to
the participants.6
The underlying issue was that the online meeting platform used to power the conven‐
tion was designed for education and not necessarily for science. While on the surface
this does not immediately pose an issue (when you consider teenagers), a built-in
“naughty-word filter” seemed to be the culprit.
The filter automatically went over all text data presented in the platform and filtered
out certain words considered “not appropriate for school.” Words that are essential
when discussing paleontology, such as, well, “bone,” are inappropriate in other set‐
tings. To top it off, even “dinosaur” was apparently inappropriate in the business con‐
text originally envisioned for the meeting platform.
6 Poppy Noor, “Overzealous Profanity Filter Bans Paleontologists from Talking About Bones”, Guardian, Octo‐
ber 16, 2020.
Why Is Data Quality a Part of a Data Governance Program?
|
121
This filtering caused a stir when one researcher found out that “Wang” was one of the
banned words, while “Johnson” was not. This issue introduced serious (unintended)
bias as both are common surnames as well as slang terms.
At the end of the day, the issue was quickly fixed, and people accepted the error in
good humor, laughing at the unintended consequences and sharing them over social
media. For the purpose of this book, however, this story presents an important lesson:
any system that acts on data (e.g., the data modification system discussed in this side‐
bar) must be developed with as full a business use case as possible in mind. Clearly,
the intended audience is part of the business context, and you obviously cannot treat
international science symposium participants like you would treat schoolchildren.
Figure 5-3 drives home the point that data quality must be tied to the specific busi‐
ness case, and it also conveys a more subtle point about data governance. Sharing
“banned word lists” between use cases also informs the use case participants of the
existence of those other cases as well as of the relevant word lists—not always a
desired scenario!
Figure 5-3. Architectures of a chat filter
122
|
Chapter 5: Improving Data Quality
Scorecard
In your organization, create a scorecard for the data sources. A useful scorecard
includes information about the origin of the data and its accuracy, completeness, and
timeliness. This scorecard will be used by data pipeline builders to make decisions
about how and where to use the data and for what purpose. More mundane informa‐
tion, such as who is the administrative owner of the data, who asked for it, and so on,
can also be useful in a scorecard.
Prioritization
First, prioritize—there are different sources of data and different uses for each source.
A data source used for prioritizing healthcare actions is very different from a data
source used to drive graphics for a lobby “heat map” display. Prioritization should be
performed with the eventual business goal in mind. Lineage can help with backtrac‐
ing the data and repurposing the origin for a different business purpose. By monitor‐
ing the lineage of data (more on that in “Lineage Tracking” on page 46) for both its
sources as well as its eventual uses—you can prioritize and expend resources on the
more critical data sources first.
Annotation
Second, annotate—make sure you have a standardized way to attach “quality infor‐
mation” to data sources. Even if you cannot provide a detailed scorecard for each
source, being able to attest that “This data has been vetted” or (just as importantly)
“This data is not vetted” is valuable, even if there are differences between the datasets
and their meaning. As you evolve your data quality program, you can further attach
more detailed information to datasets. Begin small, however, and clearly annotate the
information at hand so that it does not become “tribal knowledge” and get lost when
people in the know move on.
A common technique for annotation can be to “crowdsource” the quality information
by allowing people using the data to “vote” on or “star” data according to its quality as
they observe it through usage. This allows multiple human eyes on the data, and if
you start with a good default (data is normally untrusted) and assign a curator to
review data that is highly rated before issuing a quality signal to the rest of the organi‐
zation, you can effectively practice a fair data quality program. (This is not the only
recommendation in this chapter, though!)
Techniques for Data Quality
|
123
Cascading Problems with Tribal Knowledge
While you are certainly very familiar with the struggle of overcoming reliance on
tribal knowledge, the following is an interesting use case that further illustrates just
how problematic disregarding annotation can be.
We’ve spoken many times with a healthcare company that is going through a myriad
of issues related to lack of annotation and reliance on tribal knowledge. This company
is quickly growing and has recently taken in several new acquisitions. Part of the ben‐
efit of these acquisitions is the data that they have. This company has the vision of
being able to leverage the data from these new acquisitions along with its own data to
run some extremely powerful analytics that are sure to have prolific business impact.
The company, however, encountered an enormous snag: the enterprise dictionaries,
metadata management, and data quality management were anemic or nonexistent at
the companies they acquired. The majority of these other companies relied on tribal
knowledge, and, post-acquisition, many of the employees who had this knowledge are
no longer at the company. This has resulted in most of the acquired data (since it’s not
properly managed, and no one can know what it is without spending a lot of time and
effort to curate it) sitting in storage, just taking up space and not providing any busi‐
ness value whatsoever.
Through this issue the company has come to realize the importance of a centralized
governance strategy that it can quickly scale—even for acquisitions—so that in the
future this problem hopefully doesn’t occur again, or at least such problems can be
somewhat mitigated.
Profiling
Data profiling begins by generating a data profile: information about a range of data
values (e.g., min, max, cardinality), highlighting values that are missing and values
that are out-of-bounds (versus the average distribution) data outliers. Reviewing the
data profile enables a determination of what are considered legal values (e.g., am I
happy with the data outliers, or should I exclude those records?) and of value mean‐
ing (e.g., is a negative value in the income column an appropriate value, or should I
exclude it?).
We proceed to detail several data profiling and cleanup techniques.
Data deduplication
In a quantitative system, each record should have only one voice. However, there are
many cases in which the same record, or the same value, actually gets duplicated,
resulting in data quality issues (and potentially in increased cost). Think about a
redundant transaction system, where each transaction has an ID, and sometimes the
124
|
Chapter 5: Improving Data Quality
same transaction can appear twice (e.g., due to a transmission issue). Given the ID,
you can easily deduplicate the transactions and resolve. But think about a more chal‐
lenging use case in which you have support cases (a list of customer issues in a ticket‐
ing system), each expressed by a user input title. When writing knowledge base
articles to address these support cases, it is important to merge different user requests
that refer to the same source issue. Since a “user request” is expressed in natural lan‐
guage and can be ambiguous, this is more challenging.
Deduplicating Names and Places
Think of an entity, and it probably needs resolving or disambiguation. Two of the
most common entities that you have in datasets are names and addresses. Both of
these need to be resolved if you want to deduplicate records. For example, the same
person could be addressed as “Dr. Jill Biden,” “Jill Biden,” or “Mrs. Biden.” In all three
sets of records, it may be necessary to replace the name by a consistent identifier.
Take, for example, the impact of deduplicating author names in bibliographies, as
shown in Figure 5-4. Robert Spence is referenced as Bob Spence and as R. Spence.
Combining all these records and replacing the different variants with the canonical
version of his name greatly simplifies the set of relationships and makes deriving
insights much easier.
Figure 5-4. Deduplication can dramatically simplify the complexity of a dataset. For
example, Lisa Tweedie is also referenced as L. Tweedie. Figure adapted from a paper by
Bilgic et al., 2004.
For deduplicating names, consider using a tool such as the Google Knowledge Graph
Search API, or building one from your set of stakeholders using an open source API
such as Akutan.
Similar to names, the same place can be referred to as “the New York Stock
Exchange,” “11 Wall St,” “Wall St. & Broad Street,” “Wall and Broad,” or any of a
myriad number of combinations. If packages are noted as having been delivered to
multiple versions of this location, you might want to consolidate them into a canoni‐
cal representation of the location.
Techniques for Data Quality
|
125
Address resolution of this form is provided by the Google Places API which returns a
place ID, a textual identifier that uniquely identifies a place. Place IDs are available for
most locations, including businesses, landmarks, parks, and intersections, and will
change over time as businesses close or relocate. It can be helpful, therefore, to com‐
bine the Places API with the Google Maps Geocoding API to yield the actual location
in time. Thus, while “the New York Stock Exchange” and “11 Wall Street” yield differ‐
ent place IDs, as shown in Figure 5-5 (after all, the NYSE could relocate!), geocoding
them will return the same location.
Figure 5-5. The place IDs for “the New York Stock Exchange” and “11 Wall Street” are
different from each other but geolocate to the same location
126
| Chapter 5: Improving Data Quality
Data outliers
Another tactic is to identify outliers of the data early on and eliminate them. For
example, in a system that accepts only natural numbers (the range between 1 and
infinity, excluding fractions), such as a house number in a street address book, it
would be odd to find negative or fractional numbers, and thus it may make more
sense to delete the entire record containing the outlier value rather than manually fix‐
ing it. Discretion is advised, as (for example) for 221b Baker Street, the fictional home
of Sherlock Holmes. The UK’s Royal Mail has had to recognize “221b Baker Street,
London” as a real postal address because of all the Sherlock fans expecting it to be
genuine! But the mail to 221b redirects to the Sherlock Holmes Museum at 239 Baker
Street.
To generalize and scale: when building a dataset, make sure you can determine, for as
many fields as possible, the minimum, maximum, and expected values (fractions,
negatives, strings, zero...), and include logic to clean up records with unexpected val‐
ues (or otherwise treat them). These actions are better done early in the processing
timeline rather than later, when values have been aggregated or used in machine
learning models. It is hard to backtrack/root out issues that are discovered after data
has been processed.
Extreme values are not necessarily outliers, and care must be taken there. For exam‐
ple, a perfect SAT score is possible; however, in the US the SAT score range is 400–
1,600, and values outside this range are suspect. Look at the distribution of the values
and how the curve is shaped. The extreme edges of a “bell” curve should be treated
differently than a distribution that has two peak clusters.
Consider the example distribution in Figure 5-6 in light of the expected use case of
data. In our example, an expected bell curve with a sudden peak at the edge should
prompt a more manual investigation and understanding to rule out potential data
quality issues. Be wary of automatically dismissing such values without investigating
them, as sometimes these outliers are the result not of low-quality data, but of phe‐
nomena that should be accounted for in the use case.
Techniques for Data Quality
|
127
Figure 5-6. Example distributions of data
Lineage tracking
As mentioned earlier, lineage tracking for data is a force multiplier. If you can identify
source datasets of high quality, you can follow the use of those high-quality sources
and express an opinion of the result derivatives. In addition, if you can identify lowquality datasets, you can assume any computed derivative that is a product of one (or
more) low-quality source is also a low-quality result—a useful conclusion that should
guide the use of the data.
Furthermore, with lineage tracking you can backtrack from high-criticality use cases
and results (e.g., dashboards) and learn what data sources are feeding these. At a min‐
imum, you should prioritize the quality assessment of all sources of critical decisionsupporting outputs.
This process of monitoring quality by source should not be a one-time thing but
should be used every time a new dashboard/end product is set up. And it should be
on periodic review, because the benefit of early detection can be significant if man‐
aged correctly. The impact of bad data, which can often go unnoticed, has been dis‐
cussed above.
128
|
Chapter 5: Improving Data Quality
Data completeness
In some cases, there are data records with missing information, such as a customer
record without an address or a transaction without a tracking number. Special con‐
sideration should be given to such cases, and an informed decision must be made
whether or not to eliminate the records that are incomplete (resulting in “less data”
but more “complete” data). Alternatively, if accepting incomplete records, make sure
to include an annotation on the dataset indicating that it contains such records and
noting which fields are “OK” to be missing and the default values (if any) used as
input for those fields that are missing. This is especially important in the case of
merging datasets.
Merging datasets
During ETL processes, and in general, you should be aware of special values used in
the source datasets and make room for them during transformation/aggregation. If
one dataset uses “null” to indicate no data and another dataset uses “zero,” make sure
this information is available to future users. Ensure that the joining process equalizes
these two values into one consistent value (either null or zero; they have the same
meaning in our example). And of course, record this special new value.
Dataset source quality ranking for conflict resolution
When merging multiple datasets from different vendors, another topic comes to
mind: how to react if multiple-source datasets contain the same fields but with cer‐
tain fields having different values in them. This is a common issue in financial sys‐
tems, for example, where transactions are collected from multiple sources but
sometimes differ in the meaning of special values such as zero, negative 1. One way to
resolve that is to attach a ranking to each data source, and, in case of conflict, record
the data from the highest-ranked source.
Unexpected Sources for Data and Data Quality
Many a McDonald’s fan is disappointed when the ice cream maker in their local
branch is broken. In fact, this issue troubled Rashiq, a young McDonald’s fan in Ger‐
many, so much that he reverse-engineered the McDonald’s app, found its API, and
discovered a way to figure out whether or not the ice cream maker in a particular
branch was operational.
Rashiq tested out the code and later built a site, mcbroken.com, that reports on the
status of the ice cream makers in every branch on a global scale (Figure 5-7).
Techniques for Data Quality
|
129
Figure 5-7. Rashiq’s Twitter feed
However, there was a downside to Rashiq’s code. In order to identify whether or not
the ice cream maker was operational, Rashiq had a bot create an ice cream order and
add an ice cream to the shopping cart. If that operation was successful, Rashiq would
mark the ice cream maker in that branch as “operational.” This resulted in thousands
of half-completed orders of ice cream at McDonald’s locations, globally. Apparently,
however, the McDonald’s data team was able to control for Rashiq’s orders, as evi‐
denced by the reply from McDonald’s’ director of analytics: “I’m Lovin’ It.” McDo‐
nald’s head of communications tweeted a response (Figure 5-8).
130
|
Chapter 5: Improving Data Quality
Figure 5-8. McDonald’s responds to Rashiq’s work
Summary
Data quality is making sure that the data’s accuracy, completeness, and timeliness are
relevant to the business use case in mind. Different types of business use necessitate
different levels of the above, and you should strive to keep a scorecard of your data
sources when creating an analytics workload composed of descendants of these data
sources.
We have reviewed the importance and real-life impact of data quality, through exam‐
ples of bad data. We have discussed several techniques to improve data quality. If
there is one key piece of advice that should be taken into account from this chapter, it
is this: for data quality, handle it early, as close to the data source as possible, and
monitor resultant products of your data. When repurposing data for a different ana‐
lytics workload, revisit the sources and see if they are up to the new business task.
Summary
|
131
CHAPTER 6
Governance of Data in Flight
Data, especially data used for insights via data analytics, is a “living” medium. As data
gets collected from multiple sources, it is reshaped, transformed, and molded into
various patterns for different use cases: from a standardized “transactions table” to
allow for forecasting the next season’s business demand, to a dashboard presenting
the past yield of a new crop, and more.
Data governance should be consistent across these transformations and allow more
efficiency and frictionless security. Data governance should not introduce labor by
forcing users to register and annotate new data containers as they work to reshape
and collect data for their needs.
This chapter will discuss possible techniques and tools to enable seamless data gover‐
nance through analysis of data “in flight.”
Data Transformations
There are different ways to transform data, all of which impact governance, and we
should be aware of these before we dig in deeper. It is common to refer to these pro‐
cesses as extract-transform-load (ETL). This is a generic phrase used to indicate the
various stages of moving data between systems.
Extracting data means retrieving it from the source system in which it is stored, e.g., a
legacy DB, a file, or the results a web crawler operation. Data extraction is a separate
step, as the act of extracting data is a time-consuming retrieval process. It is advanta‐
geous to consider the extraction phase as the first step in a pipeline, allowing subse‐
quent steps to operate in batches in parallel to the continued extraction. As data is
extracted from the sources, it’s useful to perform data validation, making sure the val‐
ues retrieved are “as expected” (that the completeness of records and their accuracy
match the expected values; see Chapter 5). If you perform data validation while still
133
operating within the context of the source system, you will be unencumbered by the
different computed results performed in later stages, which may be unknown (at this
stage), and as you progress, you may lose the context of the source data. In an earlier
chapter, we discussed data preparation, which is an example of a data validation pro‐
cess. The data being extracted and validated often lands in a staging area not normally
accessible to the business users, which is where the data owners and data stewards
perform the aforementioned validation checks.
Transforming data usually involves normalization of the data: eliminating outliers,
joining from multiple sources into a single record (row), aggregating where relevant,
or even splitting a single compound column into multiple columns. Be wary of the
fact that any normalization done early, as well as any kind of general purpose clean‐
ing, is also removing information—information whose value may not have been
anticipated for a case unknown at the cleaning level. The implications are that you
should have the business context in mind when extracting data and that you may
need to revisit the source data in case the extraction process removed information
needed for a new, unforeseen use case. At this stage, if you are extracting data from a
new source, it may be worthwhile to create a scorecard for the source, describing
some of the information contexts and those that are potentially not carried over dur‐
ing transformation. See “Scorecard” on page 123 for more about scorecards.
Finally, the load process situates the data into its final destination, usually a dataanalytics-capable warehouse such as Google’s BigQuery, Snowflake, or Amazon’s
Redshift.
It is important to note that as data warehousing solutions grew to be more powerful,
the transformation process sometimes moved into the data warehousing solution,
renaming ETL as ELT.
As data undergoes transformations, it is crucial not only to keep context as expressed
by the origin but also to maintain consistency and completeness. Keeping origin con‐
text, consistency, and completeness will allow a measure of trustworthiness in the
data, which is critical for data governance.
Lineage
Following the discussion of data transformations, it is important to note the role
played by data lineage. Lineage, or provenance, is the recording of the “path” that data
takes as it travels through extract-transform-load, and other movement of data, as
new datasets and tables are created, discarded, restored, and generally used through‐
out the data life cycle. Lineage can be a visual representation of the data origins (cre‐
ation, transformation, import) and should help answer the questions “Why does this
dataset exist?” and “Where did the data come from?”
134
|
Chapter 6: Governance of Data in Flight
Why Lineage Is Useful
As data moves around your data lake, it intermingles and interacts with other data
from other sources in order to produce insight. However, metadata—the information
about the data sources and their classification—is at risk of getting lost as data travels.
For example, you can potentially ask, for a given data source, “What is the quality of
the data coming from that source?” It could be a highly reliable automatic process, or
it could be a human-curated/validated dataset. There are other sources that you
potentially trust less. As data from different sources intermingle, this information can
sometimes be lost. Sometimes it is even desirable to forgo mixing certain data sources
in order to preserve authenticity.
In addition to data quality, another common signal potentially available for source
data is sensitivity. Census information and a recently acquired client phone list are of
a certain level of sensitivity, while data scraped off a publicly available web page may
be of a different sensitivity.
Thus, data sensitivity and quality, and other information potentially available on the
origin, should filter down to the final data products.
The metadata information of the data sources (e.g., sensitivity, quality, whether or not
the data contains PII, etc.) can support decisions about whether or not to allow cer‐
tain data products to mix, whether or not to allow access to that data and to whom,
and so on. When mixing data products, you will need to keep track of where the data
came from. The business purpose for the data is of particular importance, as the cre‐
ated data products should be useful in achieving that purpose. If the business purpose
requires, for example, a certain level of accuracy for time units, make sure the lineage
of your data does not coarsen time values as the data is processed. Lineage is there‐
fore crucial for an effective data governance strategy.
How to Collect Lineage
Ideally, your data warehouse or data catalog will have a facility that starts from the
data origin and ends with whatever data products, dashboards, or models are used,
and that can collect lineage for every transaction along the way. This is far from com‐
mon, and there will likely be blindspots. You will have to either infer the information
that you need on by-products or otherwise manually curate information to close the
gap. Once you have this information, depending on how trustworthy it is, you can use
it for governance purposes.
As your data grows (and it is common for successful businesses to accumulate data at
an exponential rate) it is important to allow more and more automation into the pro‐
cess of lineage collection and to rely less and less on human curation. Automation is
important because, as we will see in this chapter, lineage can make a huge difference
in data governance, and placing human bottlenecks in the path to governance blocks
Lineage
|
135
an organization’s ability to make data accessible. Also, a broken chain of lineage (e.g.,
due to human error) can have a larger impact on derivative data products, as trust
becomes harder to come by.
Another way to collect/create lineage information is to connect to the API log for
your data warehouse. The API log is expected to contain all SQL jobs and also all pro‐
grammatic pipelines (R, Python, etc.). If there is a good job audit log, you can use it to
create a lineage graph. This allows backtracking table creation statements to the
table’s predecessors, for example. This is not as effective as just-in-time lineage
recording, as it requires backtracking the logs and batch processing. But assuming
(not always true!) that you are focusing on lineage within the data warehouse, this
method can be extremely useful.
Types of Lineage
The level of granularity of the lineage is important when discussing possible applica‐
tions for that lineage. Normally, you would want lineage at least in the table/file
level—i.e., this table is a product of this process and this other table.
In Figure 6-1, we see a very simple lineage graph of two tables joined by a SQL state‐
ment to create a third table.
Figure 6-1. Table-level lineage
A column/field-level granularity is more useful: “This table consists of the following
columns from this other table and another column from that table.” When talking
about column-level granularity, you can start to talk about specific types of data being
tracked. In structured (tabular) data, a column is normally only a single data type. An
example business case would be tracking PII: if you mark at the sources which col‐
umns are PII, you can continue tracking this PII as data is moved and new tables are
created, and you can confidently answer which tables have PII. In Figure 6-2, two col‐
umns from Table A are joined with two columns from Table B to create Table C. If
136
| Chapter 6: Governance of Data in Flight
those source columns were “known PII” (as an example), you can use lineage to
determine that Table C now contains PII as well.
Figure 6-2. Column-level lineage
Row-level lineage allows the expression of information about transactions. Datasetlevel lineage allows the expression of coarse information about data sources.
More Granular Access Controls
One of the most common use cases we have heard during our research and interviews
is that, while project/file-level access controls work, being able to more granularly
control access is highly desired. For example, column-level access control (“Provide
access to columns 1–3 and 5–8 of this table but not column 4”) allows you to lock
down, if you will, a particular column of data and still allow access to the rest of the
table.
This method works especially well for tables that contain much usable, relevant ana‐
lytics data but may also contain some sensitive data. A prime example of this would
be a telecommunications company and its retail store transactions. Each retail store’s
transaction logs not only would contain information about each item purchased,
along with date, price, etc., but also would likely contain the purchaser’s name (in the
event the purchaser is a customer of the telecom company and put the item onto their
account). An example system leveraging labels-based (or attribute-based) granular
access controls is depicted in Figure 6-3.
Lineage
|
137
Figure 6-3. Label-based security in Oracle
An analyst certainly wouldn’t need the customer’s name or account information, but
they would need the other relevant information of item purchased, location, time,
price, and so on. Instead of having to either grant full access to this table or rewrite
the table with the sensitive information removed, the sensitive column itself can have
an access control so that only certain individuals are allowed access, and all others
simply see the table with that column removed or redacted/hashed/encrypted in some
way.
As we talked about in Chapter 3, most companies do not have enough headcount to
support the constant monitoring, tagging, and rewriting of tables to remove or
restrict sensitive information. Granular access controls can enable you to get the same
result (sensitive information is guarded) while also allowing greater data
democratization.
The Fourth Dimension
As data gets created, coarsened, and discarded, it is important to realize that lineage is
a temporal state—while a certain table is currently a product of other tables, it is pos‐
sible that in a previous iteration the very same table was generated out of other data.
Looking at the “state of now” is useful for certain applications, which we will discuss
shortly, but it is important to remember that the past versions of certain data objects
are relevant to gaining a true understanding of how data is being used and accessed
138
|
Chapter 6: Governance of Data in Flight
throughout an enterprise. This requires versioning information that needs to be
accessible on the object.
How to Govern Data in Flight
Working off the assumption that lineage information is preserved and is reliable to a
certain extent, there are key governance applications that rely on lineage.
A common need that gets answered by lineage is debugging or understanding sudden
changes in data. Why has a certain dashboard stopped displaying correctly? What
data has caused a shift in the accuracy of a certain machine learning algorithm? And
so on. Finding from an end product which information feeds into that end product
and looking for changes (e.g., a sudden drop in quality, missing fields, unavailability
of certain data) could significantly speed up tasks related to end products. Under‐
standing the path and transformations of the data can also help troubleshoot data
transformation errors.
Another need that commonly requires at least field-level lineage (and said lineage
needs to be reliable) is the ability to infer data-class-level policies. For a certain col‐
umn that contains PII, I want to mark all future descendants of this column as PII
and to effect the same access/retention/masking policies of the source column to the
derivatives. You will likely need some algorithmic intelligence that allows you to effect
an identical policy if the column is copied precisely and to effect a different (or no)
policy if the column contents are nullified as a result of the creation of a derivative.
In Figure 6-4, we see a slightly more involved lineage graph. While the operations
performed to create derivatives are marked as “SQL,” note that this is not always the
case; sometimes there are other ways to transform the data (for example, different
scripting languages). As data moves from the external data source into Table D, and
as Table D is merged with Table C (the by-product of Tables A and B) and finally pre‐
sented in a dashboard, you can see how keeping lineage, especially column-level line‐
age, is important.
For example, a question asked by the business user would be “Who can I show this
dashboard to?” Let’s say that just one of the sources (Table A, Table B, the external
data source) contains PII in one of the columns, and that the desired policy on PII is
“available only to full-time employees”; if PII made it into the dashboard, that should
effect an access policy on the dashboard itself.
Lineage
|
139
Figure 6-4. Lineage workflow—if Table B contains sensitive data, that data can poten‐
tially be found in the dashboard as well
These use cases are rather specific, but broader use cases have been expressed by
CIOs and CISOs for a long time:
“Show me all the sensitive data within our data warehouse, and what systems contain
sensitive data”
This is a simple-to-express ask, and with the combination of lineage and the abil‐
ity to classify data sources, it is much simpler to answer than understanding the
multitude of data objects and systems that may be part of an organization’s data
lake.
“I want to identify certain systems as trusted, and make sure data does not exist in other,
less trusted systems without manual oversight”
This need can be addressed by enforcing egress controls on the trusted systems, a
task which is potentially simpler with a good lineage solution. In this way, you
can, for example, eliminate data being ingested from unapproved systems,
because those sources will show up in the lineage information.
“I need to report and audit all systems that process PII”
This is a common need in a post-GDPR world, and if you mark all sources of PII,
you can leverage the lineage graph to identify where PII gets processed, allowing
a new level of control.
140
|
Chapter 6: Governance of Data in Flight
Policy Management, Simulation, Monitoring, Change
Management
We already provided one example for policy management: inheritance. In essence,
data governance policies should be derived from the meaning of the data. If in a cer‐
tain organization we want to govern PII, and PII is defined as all of personal phone
numbers, email addresses, and street addresses, these individual infotypes can be
automatically detected and associated with the data class. However, scanning a table
and determining which columns contain these infotypes is expensive computation‐
ally. To correctly identify infotypes, the system will need to sample the relevant col‐
umns (without foreknowledge about which columns are sensitive, the entire table will
need to be sampled), process those through a pattern matching and/or a machine
learning model that will provide a level of confidence as to the underlying infotype,
and tag the columns appropriately.
However, this process is made much more efficient if you can just identify the event
of a column creation without changing in values from an already-tagged column.
This is where lineage comes in.
Another use for lineage information is when data change management is considered.
Let’s imagine that you want to delete a certain table; or alternatively, you want to
change the access policy on a data class; or maybe you want to set up a data-retention
control that will coarsen the data over a period of time (for example, change the data
from “GPS Coordinates” to city/state after 30 days). With lineage, you can follow the
affected data to its end products and analyze the impact of this change. Let’s say you
limit access to a certain number of fields in a table. You can now see which dash‐
boards or other systems use the data and assess the impact. A desirable result will be
highlighting the change in access to end users accessing end products, so a warning
could be issued at the time of changing the policy, alerting the administrator to the
fact certain users will lose access and potentially even allowing a level of drill down so
those users can be inspected and a more informed decision can be made.
Audit, Compliance
We often need to point at lineage information to be able to prove to an auditor or a
regulator that a certain end product (machine learning model, dashboard, etc.) was
fed into by specific and preapproved transactional information. This is necessary
because regulators will want to be able to explain the reasoning behind certain
decision-making algorithms and make sure these rely on data that was captured
according to the enterprise’s charter—for example, making sure loan approvals rely
only on specific credit information that was collected according to regulations.
Policy Management, Simulation, Monitoring, Change Management
|
141
It is increasingly common for regulated organizations to be able to prove that even
machine learning models are not biased, that decisions that are derived from data are
done so without manipulation, and that there is an unbroken “chain of trust” between
properly acquired data sources and the end-user tools (e.g., dashboards, expert sys‐
tems) that are guiding those decisions. For example, in the credit scenario just
described, a decision on a new line of credit must be traced exclusively to a list of
transactions from a trusted source, without other influences guiding the decision.
While it’s clearly incredibly important to be able to “prove” compli‐
ance in the event of an external audit, these same tools can be used
for internal auditing purposes as well. In fact, it’s best practice (as
we’ve mentioned several times in earlier chapters) to continually
assess and reassess your governance program—how it’s going,
where it may need to be modified, and how it may or may not need
to be updated as regulations and/or business needs change.
Summary
We have seen how collecting data lineage can enable data governance policies, infer‐
ence, and automation. With a lineage graph, organizations can trace data variations
and life cycle, promoting control and allowing a complete picture of the various sys‐
tems involved in data collection and manipulation. For the business user, governing
data while it is “in flight” through lineage allows a measure of trustworthiness to be
inherited from trusted sources or trusted processors in the data path. This enriched
context of lineage allows matching the right datasets (by sensitivity) to the right peo‐
ple or processes.
While lineage is essentially a technical construct, we should always keep the end busi‐
ness goal in mind. This could be “ensuring decisions are made with high-quality
data,” in which case we need to show lineage to the result from high-quality sources
(and be able to discuss the transformations along the way), or a specific business case
such as “tracing sensitive data,” as we’ve already discussed.
The technical lineage and the business lineage use cases discussed here are both
important, and we should strive to present the many technical details (e.g., intermedi‐
ate processing tables) to analysts while at the same time serving a simplified “business
view” to the business users.
142
|
Chapter 6: Governance of Data in Flight
CHAPTER 7
Data Protection
One of the key concerns of data governance is that of protecting data. Owners of data
might be concerned about the potential exposure of sensitive information to individ‐
uals or applications without authorization. Leadership might be wary of security
breaches or even of known personnel accessing data for the wrong reasons (e.g., to
check on the purchase history of a celebrity). Users of the data might be concerned
about how the data they rely on has been processed, or whether it has been tampered
with.
Data protection has to be carried out at multiple levels to provide defense in depth. It
is necessary to protect the physical data centers where the data is stored and the net‐
work infrastructure over which that traffic is carried. To do this, it is necessary to
plan how authorized personnel and applications will be authenticated and how that
authorization will be provisioned. However, it is not enough to simply secure access
to the premises and the network—there is risk involved with known personnel
accessing data that they are not supposed to be accessing, and these personnel are
inside the network. Additional forms of protection such as encryption are required so
that even if a security breach happens, the data is obfuscated.
Data protection needs to be agile because new threats and attack vectors continue to
materialize.
Planning Protection
A key aspect of data governance is determining the level of protection that needs to
be afforded to different types of data assets. Then all of the organization’s data assets
must be classified into these levels.
143
For example, at the planning level, it might be mandated that data created by the
payment-processing system needs to be secured because a malicious actor with access
to individual transaction data might be able to make spurious orders and have them
billed to the original customer. This will then have to be implemented at the authenti‐
cation level by ensuring that access to any aspect of the payments data is available only
to personnel and applications that are recognized as having a role of employee. How‐
ever, not all data is made available to all employees. Instead, the payments data might
be classified at different levels. Only aggregate data by store location, date, inventory,
and payment type might be made available to business planners. Access to individual
transactions would be authorized only to the payments support team, and even then
is granted only for specific transactions on a timed basis, when authorization is pro‐
vided in the form of a support ticket carrying the customer’s approval. Given this, it is
necessary to ensure that exfiltration of payments data is not possible. The data gover‐
nance tools and systems must support these requirements, and violations and
breaches must be captured, and alerts carried out.
Ideally, a catalog of all the data assets is created, although quite often planning, classi‐
fication, and implementation can happen at higher levels of abstraction. To carry out
the above governance policy, for example, it is not necessary that you have an explicit
catalog of all the possible aggregates of payment data—only that any aggregates that
are created are placed in a governance environment with strict boundaries.
Lineage and Quality
As discussed in Chapter 5, data lineage and quality considerations are critical aspects
of data governance. Therefore, they need to be part of the protection planning pro‐
cess. It is not enough to consider data protection of only raw data; instead, protection
of the data at each stage of its transformation needs to be considered. When aggre‐
gates are computed from protected data, those aggregates need some level of protec‐
tion that is typically equal to or less than the aggregate. When two datasets are joined,
the level of protection afforded to the joined data is often the intersection of authenti‐
cation and authorization permissions. Aggregates and joins affect the data quality,
and so governance needs to take this into account as well. If the data protection on
the raw data has restrictions on the volume of data that any person or partner is able
to access, it might be necessary to revisit what the restriction on the aggregates and
joins needs to be.
Lineage and quality checks also offer the capability to catch erroneous or malicious
use of data. Corrupt or fraudulent data may be clearer when aggregated. For example,
Benford’s Law predicts the relative incidence of leading digits on numeric values that
span multiple orders of magnitude. The last digits of numeric values are expected to
have a normal distribution. Carrying out such statistical checks is easier on aggrega‐
ted, transformed data. Once such fraud is observed, it is necessary to have the ability
144
|
Chapter 7: Data Protection
to trace back (through data lineage) where the change occurred and whether it was
the result of a data protection breach.
It is part of the data culture of an organization that data published by the organization
and used within the organization is trustworthy. Data quality, as discussed in Chap‐
ter 5, remains a fundamental goal of data protection. For this purpose, it is important
to have trust-but-verify safeguards built into data pipelines to catch quality errors
when they arise.
Level of Protection
The level of protection to be afforded to an asset should reflect the cost and likeli‐
hood of a security breach associated with that asset. This requires cataloging the types
of security breaches and the costs associated with each breach. Different levels of pro‐
tection also carry costs, and so it is necessary to identify the likelihood of a breach
occurring with the given level of protection. Then a cost analysis that balances risk
between different protection levels needs to be carried out, and a protection level
chosen.
For example, consider the raw payments-processing information stored in a data lake.
Potential security breaches might include individual transaction data being read by a
malicious actor, all the transactions within a certain time period being read, all the
transactions from a certain store being read, and so on. Similarly, there is a risk of
data being modified, corrupted, or deleted.
When considering the risk of data being modified, corrupted, or
deleted, note that this may happen on purpose due to a malicious
actor as we’ve described above, or it may happen unintentionally
because of a mistake made by an internal employee. In many of our
interviews with companies we have found that both scenarios can
and have resulted in unfavorable and at times disastrous outcomes,
and thus both warrant your attention and consideration.
The cost of a security breach at a given level of protection needs to include the cost of
that data not being available, whether because of data protection or because of a secu‐
rity breach. It is also important to realize that loss of data continuity itself can carry
costs, and so the cost of losing, say, an hour of data may affect the company’s ability to
provide accurate annual reports. The costs can also be quite indirect, such as down‐
time, legal exposure, loss of goodwill, or poor public relations. Because of all of these
factors, the costs will typically be different for different stakeholders, whether they are
end users, line-of-business decision makers, or company executives. The costs might
also accrue outside the company to customers, suppliers, and shareholders. In cases
where a granular-level cost estimate is impractical, a cost level of high/medium/low
can be assigned to guide the level of protection that is needed.
Planning Protection
|
145
There are usually a variety of choices in terms of the data protection that can be
applied. At one extreme, we might choose to not store the data at all. At the other
extreme, we might choose to make the dataset publicly available. In between are
choices such as storing only aggregated data, storing only a subset of the data, or
tokenizing certain fields. We can also choose where to store the data, perhaps driven
by regulations around geographic locations and by which roles need to be provided
access to the data. Risk levels vary between these choices, because the likelihood of a
breach, data loss, or corruption is different with each.
Classification
As covered in detail in Chapter 4, implementing data governance requires being able
to profile and classify sensitive data. This profile of the data is required to identify the
potential security breaches, their cost, and their likelihood. This in turn will allow the
data governance practitioner to select the appropriate governance policies and proce‐
dures that need to apply to the data.
There may be a number of governance directives associated with data classification,
organization, and data protection. To enable data consumers to comply with those
directives, there must be a clear definition of the categories (for organizational struc‐
ture) and classifications (for assessing data sensitivity).
Classification requires properly evaluating a data asset, including the content of its
different attributes (e.g., does a free-form text field contain phone numbers?). This
process has to take into account both the business use of the data (which parts of the
organization need to be able to access the data) and the privacy and sensitivity impli‐
cations. Each data asset can then be categorized in terms of business roles and in
terms of the different levels of data sensitivity, such as personal and private data, con‐
fidential data, and intellectual property.
Once the classification is determined and the protection level chosen by cost analysis,
the protection level is implemented through two aspects. The first aspect is the provi‐
sioning of access to available assets. This can include determining the data services
that will allow data consumers to access the data. The second aspect is prevention of
unauthorized access. This is done by defining identities, groups, and roles and assign‐
ing access rights to each.
Data Protection in the Cloud
Organizations have to rethink data protection when they move data from onpremises to the cloud or when they burst data from on-premises to the cloud for
ephemeral hybrid workloads.
146
| Chapter 7: Data Protection
Multi-Tenancy
When large enterprises that typically deploy their systems on-premises move to the
cloud, one of the biggest changes they must come to terms with is being one of many
organizations simultaneously using multi-tenant cloud architecture. This means that
it is particularly important not to leak data by leaving it in unsecured locations due to
unfounded confidence that malicious actors will not be able to authenticate into the
physical infrastructure. Whereas on-premises organizations have physical and net‐
work perimeter control, that control may be lost when going to the cloud.
Some cloud providers provide “bare metal” infrastructure or “government cloud,”
essentially providing data center management to address this change. However, rely‐
ing on such single-tenant architecture often brings increased costs, greater silos, and
technical debt.
Many of the concepts and tools of on-premises security are implemented in the
cloud. So it is possible to hew closely to the way data access, categorization, and clas‐
sification are done on premises. However, such a lift-and-shift approach can involve
giving up many of the benefits of a public cloud in terms of elasticity, democratiza‐
tion, and lower operating cost. Instead, we recommend applying cloud-native secu‐
rity policies to data that is held in the cloud, because there are better and more
modern ways to achieve data protection goals.
Use cloud identity and access management (IAM) systems rather than the Kerberosbased or directory-based authentication that you may be using on premises. This best
practice involves managing access services as well as interoperating with the cloud
provider’s IAM services by defining roles, specifying access rights, and managing and
allocating access keys for ensuring that only authorized and authenticated individuals
and systems are able to access data assets according to defined rules. There are tools
that simplify migration by providing authentication mapping during the period of
migration.
Security Surface
Another big change from on premises to cloud is the sense of vulnerability. At any
particular time, there are a number of security threats and breaches in the news, and
many of these will involve the public cloud. As we discussed in Chapter 1, much of
this is due to the increased security monitoring and auditability that public cloud sys‐
tems offer—many on-premises breaches can go unnoticed for long periods of time.
However, because of this increased media attention, organizations may be concerned
that they might be the next victim.
One of the benefits of the public cloud is the availability of dedicated, world-class
security teams. For example, data center employees undergo special screening and are
specifically trained on security. Also, a dedicated security team actively scans for
Data Protection in the Cloud
|
147
security threats using commercial and custom tools, penetration tests, quality assur‐
ance (QA) measures, and software security reviews. The security team includes
world-class researchers, and many software and hardware vulnerabilities are first dis‐
covered by these dedicated teams. At Google, for example, a full-time team known as
Project Zero aims to prevent targeted attacks by reporting bugs to software vendors
and filing them in an external database. The final argument against this sense of vul‐
nerability is that “security by obscurity” was never a good option.
The security surface is also changed by the scale and sophistication of the tools used
on the cloud. Whether it is the use of AI-enabled tools to quickly scan datasets for
sensitive data or images for unsafe content, or the ability to process petabytes of data
or to carry out processing in real time on streaming data, the cloud brings benefits in
terms of being able to apply governance practices that may not be possible on prem‐
ises. Being able to benefit from widely used and well-understood systems also reduces
the chances of employee error. Finally, using a common set of tools to be able to com‐
ply with regulatory requirements in all the places where an organization does busi‐
ness leads to a dramatically simpler governance structure within that organization.
It is therefore worth having a conversation with your cloud vendor about security
best practices, because they vary by public cloud. A public cloud where much of the
traffic is on private fiber and where data is encrypted by default will have a different
surface than one where traffic is sent over public internet.
Virtual Machine Security
A necessary part of securing your data in the public cloud is to design an architecture
that limits the effects of a security compromise on the rest of the system. Because
perimeter security is not an option, it is necessary to redesign the architecture to take
advantage of shielding and confidential computing capabilities.
For example, Google Cloud offers Shielded VM to provide verifiable integrity of
Compute Engine virtual machine (VM) instances. This is a cryptographically pro‐
tected baseline measurement of the VM’s image in order to make the virtual
machines tamperproof and provide alerts on changes in their runtime state.
This security precaution allows organizations to be confident that your instances
haven’t been compromised by boot- or kernel-level malware. It prevents the virtual
machine from being booted in a different context than it was originally deployed in—
in other words, it prevents theft of VMs through “snapshotting” or other duplication.
Microsoft Azure offers Confidential Compute to allow applications running on Azure
to keep data encrypted even when it’s in-memory. This allows organizations to keep
data secure even if someone is able to hack into the machine that is running the code.
148
|
Chapter 7: Data Protection
AWS offers Nitro Enclaves to its customers to create isolated compute environments
to further protect and securely process highly sensitive data such as PII and health‐
care, financial, and intellectual property data within Amazon EC2 instances.
Part of data governance is to establish a strong detection and response infrastructure
for data exfiltration events. Such an infrastructure will give you rapid detection of
risky or improper activity, limit the “blast radius” of the activity, and minimize the
window of opportunity for a malicious actor.
Though we are discussing “blast radius” in terms of minimizing of
the effects caused by a malicious actor, the blast radius concerns
also apply to employees mistakenly (or purposefully) sharing data
on the public internet. One concern of companies moving from
on-premises to the cloud is the increased blast radius of such leaks.
Data housed on premises, if leaked, will only leak internally and
not publicly. Companies fear that if their data is in the cloud and
leaks, it could potentially be accessible to anyone online. While this
is a valid concern, we hope that by the end of this book you will be
convinced that through the implementation and execution of a
well-thought-out governance program you can head off and pre‐
vent either scenario. Ideally, you can make your decisions about
where your data resides based on your business goals and objec‐
tives rather than on apprehension about the safety of cloud versus
on-premises data storage and warehousing.
Physical Security
Make sure that data center physical security involves a layered security model using as
many safeguards as possible, such as electronic access cards, alarms, vehicle access
barriers, perimeter fencing, metal detectors, biometrics, and laser-beam intrusion
detection. The data center should be monitored 24/7 by high-resolution interior and
exterior cameras that can detect and track intruders.
Besides these automated methods, there needs to be good old-fashioned human secu‐
rity as well. Ensure that the data center is routinely patrolled by experienced security
guards who have undergone rigorous background checks and training. Not all
employees of the company need access to the data center—so the data center access
needs to be limited to a much smaller subset of approved employees with specific
roles. Metal detectors and video surveillance need to be implemented to help make
sure no equipment leaves the data center floor without authorization.
As you get closer to the data center floor, security measures should increase, with
extra multifactor access control in every corridor that leads to the data center floor.
Physical security also needs to be concerned with what authorized personnel can do
in the data center, and whether they can access only parts of the data center. Access
Physical Security
|
149
control by sections is important in cases where regulatory compliance requires that
all maintenance be performed by citizens of a certain country/countries and/or by
personnel who have passed security clearances.
Physical security also includes ensuring an uninterrupted power supply and reducing
the chance of damage. The data center needs redundant power systems, with every
critical component having a primary and an alternate power source. Environmental
controls are also important to ensure smooth running and reduce the chances of
machine failure. Cooling systems should maintain a constant operating temperature
for servers and other hardware, reducing the risk of service outages. Fire detection
and suppression equipment is needed to prevent damage to hardware. It is necessary
to tie these systems in with the security operations console so that heat, fire, and
smoke detectors trigger audible and visible alarms in the affected zone, at security
operations consoles, and at remote monitoring desks.
Access logs, activity records, and camera footage should be made available in case an
incident occurs. A rigorous incident management process for security events is
required so as to inform customers about data breaches that may affect the confiden‐
tiality, integrity, or availability of systems or data. The US National Institute of Stand‐
ards and Technology (NIST) provides guidance on devising a security incident
management program (NIST SP 800–61). Data center staff need to be trained in for‐
ensics and handling evidence.
Physical security involves tracking the equipment that is in the data center over its
entire life cycle. The location and status of all equipment must be tracked from the
time it is acquired, through installation, and all the way to retirement and eventual
destruction. Stolen hard drives must be rendered useless by ensuring that all data is
encrypted when stored. When a hard drive is retired, the disk should be verifiably
erased by writing zeros to the drive and ensuring the drive contains no data. Malfunc‐
tioning drives that cannot be erased have to be physically destroyed in a way that pre‐
vents any sort of disk recovery—a multistage process that involves crushing,
deformation, shredding, breakage, and recycling is recommended.
Finally, it is necessary to routinely exercise all aspects of the data center security sys‐
tem, from disaster recovery to incident management. These tests should take into
consideration a variety of scenarios, including insider threats and software
vulnerabilities.
If you are using the data center in a public cloud, the public cloud provider should be
able to provide you (and any regulatory authorities) with documentation and pro‐
cesses around all of these aspects.
150
|
Chapter 7: Data Protection
Network Security
The simplest form of network security is a perimeter network security model—all
applications and personnel within the network are trusted, and all others outside the
network are not. Unfortunately, perimeter security is not sufficient for protecting sen‐
sitive data, whether on-premises or in the cloud. First of all, no perimeter is 100%
safe. At some point, some malicious actor will break into the system, and, at that
point data may become exposed. The second issue with perimeter security is that not
all applications within the network can be trusted—an application may have suffered
a security breach, or a disgruntled employee may be trying to exfiltrate data or trying
to access systems they shouldn’t have access to. Therefore, it is important to institute
additional methods of data protection to ensure that exposed data cannot be read—
this can include encrypting all stored and in-transit data, masking sensitive data, and
having processes around deleting data when it is no longer needed.
Security in Transit
Network security is made difficult because application data often must make several
journeys between devices, known as “hops,” across the public internet. The number of
hops depends on the distance between the customer’s ISP and the cloud provider’s
data center. Each additional hop introduces a new opportunity for data to be attacked
or intercepted. A cloud provider or network solution that can limit the number of
hops on the public internet and carry more of the traffic on private fiber can offer
better security than a solution that requires use of the public internet for all hops.
Google’s IP data network, for instance, consists of its own fiber, public fiber, and
undersea cables. This is what allows Google to deliver highly available and lowlatency services across the globe. Google Cloud customers can optionally choose to
have their traffic on this private network or on the public internet. Choose between
these options depending on the speed your use case requires and the sensitivity of the
data being transmitted.
Because data is vulnerable to unauthorized access as it travels, it is important to
secure data in transit by taking advantage of strong encryption. It is also necessary to
protect the endpoints from illegal request structures. One example of this is to use
gRPC (Google’s high-performance remote procedural call) as the application trans‐
port layer. gRPC is based around the idea of defining a service, specifying the meth‐
ods that can be called remotely with their parameters and return types. gRPC can use
protocol buffers as both its interface definition language (IDL) and its underlying
message interchange format. Protocol buffers are Google’s language-neutral,
platform-neutral, extensible mechanism for serializing structured data—think XML,
but smaller, faster, and simpler. Protocol buffers allow a program to introspect, that is,
to examine the type or properties of an object at runtime, ensuring that only connec‐
tions with correctly structured data will be allowed.
Physical Security
|
151
Regardless of the type of network, it is necessary to make sure applications serve only
traffic and protocols that meet security standards. This is usually done by the cloud
provider; firewalls, access control lists, and traffic analysis are used to enforce net‐
work segregation and to stop distributed denial-of-service (DDoS) attacks. Addition‐
ally, it is ideal if servers are allowed to communicate with only a controlled list of
servers (“default deny”), rather than allowing access very broadly. Logs should be rou‐
tinely examined to reveal any exploits and unusual activity.
Zoom Bombing
COVID-19 has changed how many of us work, and we have found ourselves attend‐
ing a ton of video meetings across Zoom, Google Meet, Webex, and other videocon‐
ferencing apps. It’s not surprising that this surge in popularity and usage has also
brought up many cautionary tales of “Zoom bombings,” in which bad actors use open
or poorly secured Zoom meetings to post malicious content.
First the media and then, in quick succession, regulators and Congress began to scru‐
tinize, publicize, and take legal action with respect to what were perceived as privacy
or data security flaws in the products of some technology companies. The premise
was that some of these video calls were not as private as you might think, and that
customer data was being collected and being used in ways that the customers did not
intend. The issues have been a lot more complicated because the California Con‐
sumer Privacy Act (CCPA), which took effect on January 1, 2020, instituted many
regulations pertaining to consumer data and how it can be used.
When the world changes and companies are forced to react quickly to these dynam‐
ics, things are bound to happen, and consumer privacy and data security are usually
violated. This is not new. We’ve seen this with Zoom, Facebook, YouTube, TikTok,
and many other large organizations that grow quickly and then get caught up in the
limelight. Whenever any organization rises to prominence so quickly, it emerges from
obscurity and can no longer treat privacy and data security as anything other than a
top priority.
And even with all these violations, these are still difficult conversations to have in
light of the immense contributions Zoom is making to advance the greater good by
scaling its operations so quickly to allow millions of Americans to communicate with
each other during a public health crisis.
This truly reinforces the notion that security and governance should not be an after‐
thought; they should be baked into the fabric of your organization. This chapter rein‐
forces all the elements your organization should think about when it comes to data
protection and ensuring that your data is always protected, in transit and at rest. Hap‐
pily, Zoom made several security improvements, including alerting organizers of
video calls if the link to their meeting is posted online (see Figure 7-1).
152
|
Chapter 7: Data Protection
Figure 7-1. Improvements being made after a Zoom security debacle.
Data Exfiltration
Data exfiltration is a scenario in which an authorized person or application extracts
the data they are allowed to access and then shares it with unauthorized third parties
or moves it to insecure systems. Data exfiltration can occur maliciously or acciden‐
tally. It can also happen because the authorized account has been compromised by a
malicious actor.
The traditional method of addressing data exfiltration risk relied on hardening the
physical perimeter defenses of private networks. In a public cloud, though, the net‐
work fabric is shared among multiple tenants, and there is no perimeter in the tradi‐
tional sense. Securing data in the cloud requires new security approaches and
methods of auditing data access.
It is possible to deploy specialized agents that produce telemetry about user and host
activity in virtual machines. Cloud providers also support the introduction of explicit
chokepoints, such as network proxy servers, network egress servers, and crossproject networks. These measures can reduce the risk of data exfiltration, but they
cannot eliminate it completely.
It is important to recognize common data exfiltration mechanisms, identify data at
risk of exfiltration through these vectors, and put mitigation mechanisms in place.
Table 7-1 summarizes the considerations.
Data Exfiltration |
153
Table 7-1. Data exfiltration vectors and mitigation strategies
Exfiltration vector
Use business email or mobile
devices to transmit sensitive
data from secure systems to
untrusted third parties or
insecure systems.
Download sensitive data to
unmonitored or insecure
devices.
Requisition or modify virtual
machines (VMs), deploy
code, or make requests to
cloud storage or
computation services.
Anyone with sufficient
permission can initiate
outbound transmission of
sensitive data.
Employee termination
Data at risk
Contents of organization
emails, calendars, databases,
images, planning documents,
business forecasts, and source
code.
Files of sensitive data; any
sensitive data accessed via
applications that offer a
download feature.
Mitigation mechanism
Limit volume and frequency of data transmission.
Audit email metadata such as from- and to- addresses.
Scan email content using automated tools for common
threats.
Alert on insecure channels and attempts.
Avoid storing sensitive data in files (“data lake”), and
instead keep it in managed storage such as an enterprise
data warehouse.
Establish a policy that prohibits downloads, and keep access
logs of data that is requested and served.
Regulate connections between authorized clients and cloud
services using an access security broker.
Implement dynamic watermarking in visualizations to
record the user responsible for screenshots or photographs
of sensitive information.
Add permissions-aware security and encryption on each file
using digital rights management (DRM).
Any sensitive data that is
Maintain precise, narrowly scoped permissions, and
accessible to employees in
comprehensive, immutable audit logs.
organizations’ IT departments. Maintain separate development and test datasets that
consist of simulated or tokenized data, and limit access to
production datasets.
Provide data access to service accounts with narrow
permissions, not to user credentials.
Scan all data sent to the broader internet to identify
sensitive information.
Prohibit outgoing connections to unknown addresses.
Avoid giving your VMs public IP addresses.
Disable remote management software like remote desktop
protocol (RDP).
All types of data, even
Connect logging and monitoring systems to HR software and
normally benign data like
set more conservative thresholds for alerting security teams
historical company memos, are to abnormal behavior by these users.
at risk of data exfiltration by
employees anticipating
imminent termination.a
a Michael Hanley and Joji Montelibano, “Insider Threat Control: Using Centralized Logging to Detect Data Exfiltration Near
Insider Termination”, Software Engineering Institute, Carnegie Mellon University, 2011.
In summary, because perimeter security is not an option, use of public cloud infra‐
structure requires increased vigilance and new approaches to secure data from exfil‐
tration. We recommend that organizations:
154
| Chapter 7: Data Protection
• Minimize the blast radius of data exfiltration events by compartmentalizing data
and permissions to access that data, perhaps through line of business or by com‐
mon workloads that access that data.
• Use fine-grained access control lists and grant access to sensitive data sparingly
and in a time-bound manner.
• Provide only simulated or tokenized data to development teams, because the
ability to create cloud infrastructure creates special risks.
• Use immutable logging trails to increase transparency into the access and move‐
ment of data in your organization.
• Restrict and monitor ingress and egress to machines in your organization using
networking rules, IAM, and bastion hosts.
• Create a baseline of normal data flows, such as amounts of data accessed or trans‐
ferred, and geographical locations of access against which to compare abnormal
behaviors.
Virtual Private Cloud Service Controls (VPC-SC)
Virtual Private Cloud Service Controls (VPC-SC) improves the ability of an organiza‐
tion to mitigate the risk of data exfiltration from cloud-native data lakes and enter‐
prise data warehouses. With VPC-SC, organizations create perimeters that protect
the resources and data of an explicitly specified set of services (Figure 7-2).
Figure 7-2. VPC-SC creates a perimeter around a set of services and data so that the
data is inaccessible from outside the perimeter, even to personnel and applications with
valid credentials.
Data Exfiltration |
155
VPC-SC thus extends private access and perimeter security to cloud services and
allows unfettered access to data for applications within the perimeter without open‐
ing up that data to access from malicious insiders, compromised code, or malicious
actors with stolen authentication credentials. Resources within a perimeter can be
accessed only from clients within authorized VPC networks (either the public cloud
network or an explicitly allowed on-premises one). It is possible to restrict internet
access to resources within a perimeter through allowed IP addresses or ranges of
addresses.
Clients within a perimeter that have private access to resources do not have access to
unauthorized (potentially public) resources outside the perimeter—this is what limits
the data exfiltration risk. Data cannot be copied to unauthorized resources outside
the perimeter.
VPC-SC, when used in conjunction with a restricted virtual IP, can be used to prevent
access from a trusted network to storage services that are not integrated with VPC
service controls. The restricted VIP (virtual IP) also allows requests to be made to
services supported by VPC Service Controls without exposing those requests to the
internet. VPC Service Controls provides an additional layer of security by denying
access from unauthorized networks, even if the data is exposed by misconfigured
Cloud IAM policies.
We recommend that you use the dry-run features of VPC-SC to monitor requests to
protected services without preventing access. This will enable you to monitor
requests to gain a better understanding of request traffic to your projects and provide
a way to create honeypot perimeters to identify unexpected or malicious attempts to
probe accessible services.
Note that VPC-SC limits the movement of data, rather than metadata, across a ser‐
vice perimeter via supported services. While in many cases VPC-SC also controls
access to metadata, there may be scenarios in which metadata can be copied and
accessed without VPC-SC policy checks. It is necessary to rely on Cloud IAM to
ensure appropriate control over access to metadata.
Secure Code
Data lineage is of no effect if the application code that produces or transforms the
data is not trusted. Binary authorization mechanisms such as Kritis provide software
supply-chain security when deploying container-based applications. The idea is to
extend the Kubernetes-managed runtime and enforce security policies at deploy time.
In Google Cloud, binary authorization works with container images from Container
Registry or another container image registry and extends Google Kubernetes Engine
(GKE). This allows the organization to scan built containers for vulnerabilities before
deploying to systems that can access sensitive data.
156
|
Chapter 7: Data Protection
Binary authorization implements a policy model, where a policy is a set of rules that
governs the deployment of container images to an operational cluster. Rules in a pol‐
icy specify the criteria that an image must pass before it can be deployed. A typical
policy requires a container image to have a verified digital signature before it is
deployed.
In this type of policy, a rule specifies which trusted authorities, called signers, must
assert that required processes have completed and that an image is ready to move to
the next stage of deployment. A signer may be a human user or more often it’s a
machine process such as a build-and-test system or a part of your continuous deploy‐
ment pipeline.
During the development life cycle, signers digitally sign globally unique container
image descriptors, thereby creating certificated statements called attestations. Later,
during the deploy phase, Binary Authorization uses attestors to verify the certificate
indicating that required processes in your pipeline have been completed.
Zero-Trust Model
A perimeter-based security model is problematic, because when that perimeter is
breached, an attacker has relatively easy access to a company’s privileged intranet. As
companies adopt mobile and cloud technologies, the perimeter becomes increasingly
difficult to enforce. By shifting access controls from the network perimeter to indi‐
vidual users and devices, a zero-trust security model allows users to access enterprise
applications from virtually any location without the need for a traditional VPN.
The zero-trust model assumes that an internal network is untrustworthy and builds
enterprise applications based on the assumption that all network traffic emanates
from a zero-trust network, i.e., the public internet. Instead of relying on the IP
address of the origin, access to an application depends solely on device and user cre‐
dentials. All access to enterprise resources is authenticated, authorized, and encrypted
based on device state and user credentials. Fine-grained access to different parts of
enterprise resources is enforced, and the goal is to make the user experience of
accessing enterprise resources effectively identical between internal and public
networks.
The zero-trust model consists of a few specific parts:
• Only a device that is procured and actively managed by the enterprise is allowed
to access corporate applications.
• All managed devices need to be uniquely identified using a device certificate that
references the record in a device inventory database, which needs to be
maintained.
Data Exfiltration |
157
• All users are tracked and managed in a user database and a group database that
tightly integrate with HR processes that manage job categorization, usernames,
and group memberships for all users.
• A centralized user-authentication portal validates two-factor credentials for users
requesting access to enterprise resources.
• An unprivileged network that very closely resembles an external network,
although within a private address space, is defined and deployed. The unprivi‐
leged network only connects to the internet, and to limited infrastructure and
configuration management systems. All managed devices are assigned to this
network while physically located in the office, and there needs to be a strictly
managed access control list (ACL) between this network and other parts of the
network.
• Enterprise applications are exposed via an Internet-facing access proxy that
enforces encryption between the client and the application.
• An access control manager interrogates multiple data sources to determine the
level of access given to a single user and/or a single device at any point in time.
The latter is called endpoint verification.
Identity and Access Management
Access control encompasses authentication, authorization, and auditing. Authentica‐
tion determines who you are, authorization determines what you can do, and audit‐
ing logs record what you did.
Authentication
The identity specifies who has access. This could be an end user who is identified by a
username, or an application identified by a service account.
User accounts represent a data scientist or business analyst, or an administrator. They
are intended for scenarios (e.g., notebooks, dashboard tools, or administration tools)
in which an application needs to access resources interactively on behalf of a human
user.
Service accounts are managed by Cloud IAM and represent nonhuman users. They
are intended for scenarios in which an application needs to access resources automat‐
ically. Service accounts are essentially robot accounts that are defined to have a subset
of the permissions held by the creator of the service account. Typically, they are cre‐
ated to embody the limited set of permissions required by applications that are run
on the account creator’s behalf.
158
|
Chapter 7: Data Protection
Cloud APIs reject requests that do not contain a valid application credential (no
anonymous requests are processed). Application credentials need to provide the
required information about the caller making the request. Valid credential types are:
API keys
Note that an API key says only that this is a registered application. If the applica‐
tion requires user-specific data, the user needs to authenticate themselves as well.
This can be difficult, and so API keys are often used only to access services (e.g.,
requests for stock market quotes) that do not need user credentials but just a way
to ensure that this is a paid subscriber.
Access tokens such as OAuth 2.0 client credentials
This credential type involves two-factor authentication and is the recommended
approach for authenticating interactive users. The end user allows the application
to access data on their behalf.
Service account keys
Service accounts provide both the application credential and an identity.
Authorization
The role determines what access is allowed to the identity in question. A role consists
of a set of permissions. It is possible to create a custom role to provide granular access
to a custom list of permissions. For example, the role roles/bigquery.metadata
Viewer, when assigned to an identity, allows that person to access metadata (and only
the metadata, not the table data) of a BigQuery dataset.
To grant multiple roles to allow a particular task, create a group, grant the roles to
that group, and then add users or other groups to that group. You might find it help‐
ful to create groups for different job functions within your organization and to give
everyone in those groups a set of predefined roles. For example, all members of your
data science team might be given dataViewer and jobUser permissions on data ware‐
housing datasets. This way, if people change jobs, you’ll just need to update their
membership in the appropriate groups instead of updating their access to datasets
and projects one dataset or project at a time.
Another reason to create a custom role is to subtract permissions from the predefined
roles. For example, the predefined role dataEditor allows the possessor to create,
modify, and delete tables. However, you might want to allow your data suppliers to
create tables but not to modify or delete any existing tables. In such a case, you would
create a new role named dataSupplier and provide it with the specific list of
permissions.
Identity and Access Management
|
159
Normally, access to resources is managed individually, resource by resource. An iden‐
tity does not get the dataViewer role or the tables.getData permission on all
resources in a project; rather, the permission should be granted on specific datasets or
tables. As much as possible, avoid permission/role creep; err on the side of providing
the least amount of privileges to identities. This includes restricting both the roles
and the resources on which they are provided. Balance this against the burden of
updating permissions on new resources as they are created. One reasonable compro‐
mise is to set trust boundaries that map projects to your organizational structure and
set roles at the project level—IAM policies can then propagate down from projects to
resources within the project, thus automatically applying to new datasets in the
project. The problem with such individualized access is that it can quickly get out of
hand.
Another option for implementing authorization is to use Identity-Aware Proxy (IAP).
IAP lets you establish a central authorization layer for applications accessed by
HTTPS, so you can use an application-level access control model instead of relying
on network-level firewalls. IAP policies scale across your organization. You can define
access policies centrally and apply them to all of your applications and resources.
When you assign a dedicated team to create and enforce policies, you protect your
project from incorrect policy definition or implementation in any application. Use
IAP when you want to enforce access control policies for applications and resources.
With IAP, you can set up group-based application access: a resource could be accessi‐
ble for employees and inaccessible for contractors, or accessible only to a specific
department.
Policies
Policies are rules or guardrails that enable your developers to move quickly, but
within the boundaries of security and compliance. There are policies that apply to
users—authentication and security policies, such as two-factor authentication, or
authorization policies that determine who can do what on which resources—and
there are also policies that apply to resources and are valid for all users.
Where possible, define policies hierarchically. Hierarchical policies allow you to cre‐
ate and enforce a consistent policy across your organization. You can assign hierarch‐
ical policies to the organization as a whole or to individual business units, projects, or
teams. These policies contain rules that can explicitly deny or allow roles. Lower-level
rules cannot override a rule from a higher place in the resource hierarchy. This allows
organization-wide admins to manage critical rules in one place. Some hierarchical
policy mechanisms allow the ability to delegate evaluation of a rule to lower levels.
Monitor the use of the rules to see which one is being applied to a specific network or
data resource. This can help with compliance.
160
| Chapter 7: Data Protection
Context-Aware Access is an approach that works with the zero-trust network security
model to enforce granular access control based on a user’s identity and the context of
the request. Use Context-Aware Access when you want to establish fine-grained
access control based on a wide range of attributes and conditions, including what
device is being used and from what IP address. Making your corporate resources
context-aware improves your security posture. For example, depending on the policy
configuration, it is possible to provide edit access to an employee using a managed
device from the corporate network, but provide them only view access to the data if
the access is from a device that has not been fully patched.
Instead of securing your resources at the network level, Context-Aware Access puts
controls at the level of individual devices and users. Context-Aware Access works by
leveraging four key pieces of technology that have all been discussed in this chapter:
Identity-Aware Proxy (IAP)
A service that enables employees to access corporate apps and resources from
untrusted networks without the use of a VPN.
Cloud identity and access management (Cloud IAM)
A service that manages permissions for cloud resources
Access context manager
A rules engine that enables fine-grained access control.
Endpoint verification
A way of collecting user device details.
Data Loss Prevention
In some cases, especially when you have free-form text or images (such as with
records of support conversations), you might not even know where sensitive data
exists. It is possible that the customer has revealed their home address or credit card
number. It can therefore be helpful to scan data stores looking for known patterns
such as credit card numbers, company confidential project codes, and medical infor‐
mation. The result of a scan can be used as a first step toward ensuring that such sen‐
sitive data is properly secured and managed, thus reducing the risk of exposing
sensitive details. It can also be important to carry out such scans periodically to keep
up with growth in data and changes in use.
AI methods such as Cloud Data Loss Prevention can be used to scan tables and files
in order to protect your sensitive data (see Figure 7-3). These tools come with built-in
information type detectors to identify patterns, formats, and checksums. They may
also provide the ability to define custom information type detectors using dictionar‐
ies, regular expressions, and contextual elements. Use the tool to de-identify your
Identity and Access Management
|
161
data, including masking, tokenization, pseudonymization, date shifting, and more, all
without replicating customer data.
Figure 7-3. Scanning a BigQuery table using Cloud DLP.
To redact or otherwise de-identify sensitive data that the Cloud DLP scan found, pro‐
tect the data through encryption, as discussed in the next section.
Encryption
Encryption helps to ensure that if the data accidentally falls into an attacker’s hands,
they cannot access the data without also having access to the encryption keys. Even if
an attacker obtains the storage devices containing your data, they won’t be able to
understand or decrypt it. Encryption also acts as a “chokepoint”—centrally managed
encryption keys create a single place where access to data is enforced and can be
audited. Finally, encryption contributes to the privacy of customer data—it allows
systems to manipulate data, for example, for backup, and it allows engineers to sup‐
port the infrastructure without providing access to content.
It is possible to use a public cloud provider’s encryption-at-rest and encryption-intransit mechanisms to ensure that low-level infrastructure access (to hard drives or
network traffic, for example) does not afford the ability to read your data. Sometimes,
162
| Chapter 7: Data Protection
however, regulatory compliance might require you to make sure that your data is
encrypted with your own keys. In such cases, you can use customer-managed encryp‐
tion keys (CMEK), and you can manage your keys in a key management system
(KMS), even in Cloud KMS, GCP’s central key management service. Then you can
designate datasets or tables that you want to be encrypted using those keys.
Multiple layers of key wrapping are used so that the master keys aren’t exposed out‐
side of KMS (see Figure 7-4). Every CMEK-protected table has a wrapped key as part
of the table metadata. When the native cloud tools access the table, they send a
request to Cloud KMS to unwrap the key. The unwrapped table key is then used to
unwrap separate keys for each record or file. There are a number of advantages to this
key-wrapping protocol that reduce the risk should an unwrapped key be leaked. If
you have an unwrapped file key, you can’t read any other files. If you have an unwrap‐
ped table key, you can only unwrap file keys after you pass access control checks. And
Cloud KMS never discloses the master key. If you delete the key from KMS, the other
keys can never be unwrapped.
Figure 7-4. Envelope encryption with data encryption key (DEK) and key encryption key
(KEK). The KEKs are managed centrally in a KMS, which rotates keys through the use
of a key ring.
A common use of encryption is to delete all records associated with a specific user,
often in response to a legal request. It is possible to plan for such deletion by assign‐
ing a unique encryption key to each userId and encrypting all sensitive data corre‐
sponding to a user with that encryption key. In addition to maintaining user privacy,
you can remove the records for a user simply by deleting the encryption key. This
approach has the advantage of immediately making the user records unusable (as
long as the deleted key is not recoverable) in all the tables in your data warehouse,
including backups and temporary tables. To query the encrypted table, you need to
decrypt the data before querying. This approach is called crypto-shredding.
Identity and Access Management
|
163
Differential Privacy
Another concept for keeping data secure, especially when discussing private data, is
differential privacy. Differential privacy is necessary when you want to share a dataset
that contains highly personal or otherwise sensitive information without exposing
any of the involved parties to identification. This means describing aggregate data
while withholding information about the individuals. However, this is not a simple
task—there are ways to statistically re-identify individuals in the dataset by crossmatching different dimensions of aggregate data (and knowing something about the
individual). For example, you can extract a particular salary from an aggregate aver‐
age—if you run multiple averages across the salary recipient’s home neighborhood,
their age group, and so on, you will eventually be able to compute the salary.
There are common techniques for ensuring differential privacy:
k-anonymity
k-anonymity means that aggregates returned from queries to the datasets are rep‐
resenting groups of at least k individuals, or are otherwise expanded to include k
individuals (Figure 7-5). The value of k is determined by the size of the dataset
and other considerations relevant to the particular data represented.
Figure 7-5. Example of a k-anonymized table in which the age was replaced with a kanonymized value.
Adding “statistically insignificant noise” to the dataset
For fields such as age or gender where there is a discrete list of possible values,
you can add statistical noise to the dataset so that aggregates are skewed slightly
to preserve privacy but the data remains useful. Examples of such techniques that
generalize the data and reduce granularity are l-diversity and t-distance.
164
| Chapter 7: Data Protection
Access Transparency
It is important for safeguarding access to the data that any access to the data is trans‐
parent. Only a small number of on-call engineers should be able to access user data in
the production system, and even then it should only be to ensure safe running of the
system. Whenever someone in the IT department or cloud provider accesses your
data, you should be notified.
In Google Cloud, for example, Access Transparency provides you with logs that cap‐
ture the actions Google personnel take when accessing your content. Cloud Audit
Logs helps you answer questions about “who did what, where, and when?” in your
cloud projects. While Cloud Audit Logs provides these logs about the actions taken
by members within your own organization, Access Transparency provides logs of the
actions taken by personnel who work for the cloud provider.
You might need Access Transparency logs data for the following reasons:
• To verify that cloud provider personnel are accessing your content only for valid
business reasons, such as fixing an outage or in response to a support request
• To verify and track compliance with legal or regulatory obligations
• To collect and analyze tracked access events through an automated security infor‐
mation and event management (SIEM) tool
Note that the Access Transparency logs have to be used in conjunction with Cloud
Audit Logs because the Access Transparency logs do not include access that origi‐
nates from a standard workload allowed through Cloud IAM policies.
Keeping Data Protection Agile
Data protection cannot be rigid and unchanging. Instead, it has to be agile to take
account of changes in business processes and respond to observed new threats.
Security Health Analytics
It is important to continually monitor the set of permissions given to users to see
whether any of those are unnecessarily broad. For example, we can periodically scan
user-role combinations to find how many of the granted permissions are being used.
In Figure 7-6, the second user has been granted a BigQuery Admin role but is using
only 5 of the 31 permissions that the role grants. In such cases, it is better to either
reduce the role or create a more granular custom role.
Keeping Data Protection Agile
|
165
Figure 7-6. Scan the use of permissions by users to narrow down role definitions or pro‐
vide the users more fine-grained access.
Data Lineage
A key attribute of keeping data protection agile is to understand the lineage of every
piece of data. Where did it come from? When was it ingested? What transformations
have been carried out? Who carried out these transformations? Were there any errors
that resulted in records being skipped?
It is important to ensure that data fusion and transformation tools (see Figure 7-7)
provide such lineage information, and that this linear information is used to analyze
errors being encountered by production systems.
Figure 7-7. It is important to maintain the data lineage of all enterprise data and
address errors in ingest processes.
166
| Chapter 7: Data Protection
Event Threat Detection
The overall security health needs to be continually monitored as well. Network secu‐
rity logs need to be analyzed to find the most frequent causes of security incidents.
Are a number of users trying (and failing) to access a specific file or table? It is possi‐
ble that the metadata about the file or table has been breached. It is worth searching
for the source of the metadata leak and plugging it. It is also advisable to secure the
table before one of the attacks succeeds.
Instead of limiting ourselves to scanning security logs, it is worth modeling network
traffic and looking for anomalous activity in order to identify suspicious behavior
and uncover threats. For example, an unusual number of SSH logins by an authorized
employee might be a sign of data exfiltration, and the definition of “unusual” in this
case can be learned by an AI model that compares the activity of an employee against
their peers who are doing a similar role and working on similar projects.
Data Protection Best Practices
While data protection is top of mind for all industries and users, there are different
approaches from one industry to another, including some stricter approaches in
industries such as healthcare, government, and financial services that routinely deal
in sensitive data.
The healthcare industry has been dealing with medical records for decades. As
healthcare providers have moved more and more to digital tools for record keeping,
the industry has experienced globally known incidents. The ransomware cyberattack
that affected more than 60 trusts within the United Kingdom’s National Health Ser‐
vice (NHS) spread to more than 200,000 computer systems in 150 countries, and the
list continues to grow.1
Similarly, financial institutions have been dealing with cyberattacks. We still remem‐
ber when Equifax announced a massive cybersecurity breach in 2017 in which cyber‐
criminals accessed the personal data of some 145.5 million Equifax customers,
including full names, social security numbers, birth dates, addresses, and driver’s
license numbers. At least 209,000 customers’ credit card credentials were taken in the
attack.2 Since then, a number of more visible data breaches have taken place, and
many if not all of them affected managed systems that were part of the companies’
own data centers.
1 Roger Collier, “NHS Ransomware Attack Spreads Worldwide”, Canadian Medical Association Journal 189, no.
22 (June 2017): E786–E787.
2 David Floyd, “Was I Hacked? Find Out If the Equifax Breach Affects You”, Investopedia, updated June 25,
2019.
Data Protection Best Practices
|
167
The natural question that comes to mind is: why are these data breaches still taking
place. even with the effective processes and tools that are available, and despite a con‐
certed focus to protect data? It comes down to how the best practices are imple‐
mented, and whether the institutions are staying on top of their data governance
processes day in and day out, 24/7, with no complacency. In the previous sections, we
have highlighted a comprehensive way of protecting the data; in cloud, with physical
and network security, and with advanced IAM.
Professionals in each industry—healthcare, financial institutions, retail, and others—
are trying to establish what they believe to be the best practices for their world when
it comes to data protection. As an example of those best practices, let’s start with what
data protection experts in the healthcare industry are suggesting to their users.
A data breach in healthcare can occur in several forms. It could be a criminal cyberat‐
tack to access protected health data for the purpose of committing medical identity
theft, or it could be an instance of a healthcare employee viewing patient records
without authorization.
Organizations in the healthcare industry have to be very diligent in protecting sensi‐
tive patient, financial, and other types of datasets, and must stay on top of this 24/7,
throughout their entire operations, by educating their employees and by utilizing
best-in-class security tools and best practices for the industry. Here are some of the
recommended best practices for the healthcare industry.
Separated Network Designs
Hackers utilize various methods to gain access to healthcare organizations’ networks.
IT departments in the healthcare industry should rigorously deploy tools such as fire‐
walls and antivirus and anti-malware software rigorously. However, focusing on
perimeter security is not enough. Healthcare firms should adopt network design
approaches that separate networks so that intruders cannot access the patient data
even if they are able to obtain access to parts of their network. Some firms are already
practicing and benefitting from these practices at the network-design level.
Physical Security
Paper-based record keeping of much of the sensitive data is still a very common prac‐
tice in the healthcare industry, and it is imperative that healthcare providers provide
physical security with locked cabinets and doors, cameras, and so on, while physically
securing the IT equipment where sensitive data is stored, including providing cable
locks to the laptops within offices.
168
|
Chapter 7: Data Protection
Physical Data Breach in Texas
Throughout this book we have discussed (and will discuss further) security around
your data in on-premises storage systems and/or in cloud-based storage. There are
use cases, however, around securing actual physical data (like papers or tapes), as we
pointed out in the introduction to this section.
While it’s a bit of an older example, an event that occurred in Texas in 2011 illustrates
the importance of physical security.
A data contractor was working for Science Applications International Corporation
(SAIC), the firm that handles data for TRICARE, the federal government and military
healthcare program. In his car, the contractor had the backup tapes of electronic
healthcare records for over 4.6 million active and retired military personnel. All of
these tapes were stolen from the car, compromising the health information of not
only the active and retired military personnel, but also their families.
SAIC clarified in a press release that while the tapes medical record data such as social
security numbers, diagnoses, and lab reports), financial data was not included on the
tapes and thus was not leaked.
SAIC and TRICARE set up incident response call centers to help patients deal with
the security breach and to help them to place a fraud alert on their credit reports if
needed, but they also made a statement:
If you are a citizen in the modern society, if you have a credit card, if you shop
online, if you have information stored, you should anticipate that some day your
information will get stolen.3
This statement isn’t wholly inaccurate even a decade later, but what it fails to consider
is the role of governance in preventing these sorts of things from happening in the
first place—which is likely to be the main reason you are reading this book.
You should be aware of and consider implementing additional security measures for
any physical data that you may have. Healthcare, as in this example, is an obvious sce‐
nario, but there are many instances in which there may be physical sensitive data that
needs to be treated and/or needs to be something that you educate your employees on
and build into your data culture (see Chapter 9 for more on this).
As in this example, even the methods by which you transport physical data from one
location to another should be considered. Often we think about how data is transfer‐
red between applications, where it’s stored, whether or not it’s encrypted, and so on.
But as this example shows, you should also consider how you are physically trans‐
porting data, if this is something you or your company will be doing. Will it be by car?
What are the safeguards that are in place to ensure that data is always being watched
3 Jim Forsyth, “Records of 4.9 Million Stolen from Car in Texas Data Breach”, Reuters, September 29, 2011.
Data Protection Best Practices
|
169
over or protected? Do you have any safeguards in place in the event of this data being
stolen? Is it unreadable? Is it encrypted?
Hopefully, if you note these areas of consideration and generate a plan for how to
treat and protect your physical data, what happened to SAIC and TRICARE in 2011
won’t happen to you.
Portable Device Encryption and Policy
Data breaches in the healthcare industry in recent years have occurred largely
because a laptop or storage device containing protected health information was lost
or stolen. A key measure that healthcare organizations should always undertake to
prevent a data breach due to a stolen portable device is to encrypt all devices that
might hold patient data, including laptops, smartphones, tablets, and portable USB
drives. Also, in addition to providing encrypted devices for their employees, health‐
care organizations should establish a strong policy against carrying data on an unen‐
crypted personal device. We are seeing more and more bring-your-own-device
(BYOD) policies adopted by various institutions, and many healthcare providers are
now using mobile device management (MDM) software to enforce those policies.
Data Deletion Process
A key lesson that data-breach victims have learned is the need for a data-deletion pol‐
icy, because as more data is held by an organization, there is more for intruders to
steal. Healthcare institutions should deploy a policy mandating the deletion of patient
and other information that’s no longer needed, while complying with regulations that
require records to be kept for a certain duration. Furthermore, regular audits must be
exercised to ensure that policies are followed and that organizations know what data
is where, what might be deleted, and when it can be deleted.
Electronic medical device and OS software upgrades
One of the areas that healthcare providers and their IT organizations need to pay
closer attention to is medical device software and OS upgrades. Intruders have dis‐
covered that healthcare providers were not always so diligent and still utilize outdated
OS-based medical devices that become easy targets to hack, despite update recom‐
mendations from the healthcare device vendors’. While these updates may seem dis‐
ruptive to the providers and their employees, a data breach is much worse for a
provider. Keeping these devices patched and up to date will minimize these
vulnerabilities.
170
|
Chapter 7: Data Protection
Data breach readiness
There is no such thing as preventing every possible IT security incident; this is exactly
why institutions should have a well-established plan to deploy when and if a data
breach occurs. Educating employees on what a HIPAA violation is, how to avoid
phishing, avoiding target attacks, and choosing strong passwords are among a few
simple steps that healthcare institutions can take.
The healthcare industry best practices and suggestions we’ve highlighted apply to
other industries as well. Institutions at large should implement processes and proce‐
dures to reduce in-the-moment thinking when it comes to dealing with data
breaches. Automation and well-planned and commented response are key in dealing
with a potential data breach. Establishing internal data security controls will reduce
the risk of data breaches while improving regulatory compliance. In summary, organ‐
izations should establish the following steps in their company-wide process:
• Identify systems and data that need to be protected.
• Continuously evaluate possible internal and external threats.
• Establish data security measures to manage identified risks.
• Educate and train employees regularly.
• Monitor, test, and revise—and remember that the IT world risks change all the
time.
Why Do Hackers Target the Healthcare Industry?
Every year, we see study after study on the impact of security breaches in the health‐
care industry, where hackers demand large ransoms from healthcare vendors and
providers. 2019 was no different; in fact, according to a study by Emsisoft, it has
reached record levels, costing healthcare industry members over $7.5 billion just in
the US, where over 100 state and local government agencies, over 750 providers,
nearly 90 universities, and more than 1,200 schools have been affected.4 The results
were not just an inconvenience of expense but a massive disruption to healthcare
delivery: surgeries were postponed, and in some cases patients had to be transferred
to other hospitals to receive the urgent care they needed—not to mention the disrup‐
tion to payment systems and collection systems. Even student grades were lost. 2020
was no better in terms of healthcare data breaches, according to HealthITSecurity.
The World Privacy Forum website has an interactive map that is a great resource for
getting the latest on medical data breaches (see Figure 7-8).
4 Emsisoft Malware Lab, “The State of Ransomware in the US: Report and Statistics 2019”, December 12, 2019.
Data Protection Best Practices
|
171
Figure 7-8. An example from the World Privacy Forum website shows medical data
breaches in the state of Kentucky in 2018.
Many of us regularly wonder, “Why is this happening, and why is this happening
especially in the healthcare industry?” We are alarmed to discover that a large
percentage of these institutions did not even know how often these breaches were
happening, and only a small fraction had the processes in place to recover and protect
their data. It is especially worrying to know that many of the breaches were happen‐
ing via very smart and digitally connected IoT devices. The industry has to do much
better going forward.
172
|
Chapter 7: Data Protection
Hackers target the healthcare industry largely because it is one of the few industries
with legacy operating systems like Microsoft XP or Windows 7 in their medical
devices, and it has not been industry practice to stay on top of OS patches when they
become available. In fact, some of these older OS-based medical devices will no
longer be supported by Microsoft, meaning the vendors need to ensure their users
upgrade to a device that is up to date and better prepared for today’s security
breaches. Failing to patch or renew their devices will continue to increase the risk.
Manufacturing a medical device is not easy; it takes many years to build one, and then
it is in use for 15 years or longer. So these devices become easy targets during their
20+ years of use, not to mention that many of them are shipped to other countries for
secondary and tertiary use. Considering how often the electronics and software in
our daily lives change (e.g., cell phones have an average life cycle of two years before
we refresh them), you can see why hackers target healthcare devices.
The healthcare industry must be much more proactive in its approach to data protec‐
tion. It must utilize best practices around data protection; have total control of assets;
continuously assess gaps and processes; leverage best practices and experts from
NIST, HITRUST, and others in the cybersecurity industry; and ensure that appropri‐
ate budgets are allocated for upgrading legacy medical device software and endpoints
that are open to hackers. A single weak point is enough for a hacker to impact an
entire network.
Summary
In this chapter, we looked at what is perhaps the key concern of data governance—
protecting data. A necessary part of planning protection is ensuring that lineage is
tracked and quality monitored so that the data remains trustworthy. The planning
process should identify the levels of protection and decide the kind of protection
afforded at each level. Then the classification process itself—classifying data assets
into levels—needs to be planned for.
A key concern with data stored in the public cloud is multi-tenancy. There could be
other workloads running on the same machines that are carrying out data processing
for your company. Therefore, it is important to consider the security surface and take
into account virtual machine security. Typically, physical security is provided by the
cloud provider, and your responsibility for network security is limited to getting the
data to/from the cloud (security in transit) and configuring the appropriate security
controls, such as VPC-SC. Important aspects to consider are data exfiltration, zero
trust, and how to set up Cloud IAM. These involve setting up governance of authenti‐
cation, authorization, and access policies.
Because your data may be used by data science users who need access to all of the
data, Cloud IAM and row-based security may be insufficient. You must also deter‐
mine the need for the masking, tokenization, or anonymization of sensitive
Summary
|
173
information that may not be needed to train machine learning models. A clear under‐
standing of what to encrypt and how to implement differential privacy, comprehen‐
sive access monitoring, and the ablity to change security profiles in response to
changing risks all become more important the more the data gets used for machine
learning.
The data protection governance process should also create policies on when separate
network designs are necessary, how portable devices have to be treated, when to
delete data, and what to do in the case of a data breach. Finally, data protection needs
to be agile, because new threats and attack vectors continue to materialize. So the
entire set of policies should be periodically revisited and fine-tuned.
174
|
Chapter 7: Data Protection
CHAPTER 8
Monitoring
In previous chapters, we explained what governance is; discussed the tools, people,
and processes of governance; looked at data governance over a data life cycle; and
even did a deeper dive into governance concepts, including data quality and data
protection.
In this chapter, we will do a deep dive into monitoring as a way to understand how
your governance implementations are performing day-to-day, and even on a longerterm basis. You will learn what monitoring is, why it is important, what to look for in
a monitoring system, what components of governance to monitor, the benefits of
monitoring, and the implications of doing so. You have implemented governance in
your organization, so how do you know what’s working and what’s not? It’s important
to be able to monitor and track the performance of your governance initiatives so you
can report to all the stakeholders the impact the program has had on the organiza‐
tion. This allows you to ask for additional resources, course-correct if/as needed,
learn from the wins and failures, and really showcase the impact of the governance
program—not to mention making the most of potential growth opportunities that
might become more visible to the chief data and digital officers of your organization.
So what is monitoring? Let’s start by introducing the concept.
What Is Monitoring?
Monitoring allows you to know what is happening as soon as it happens so you can
act quickly. We’re in a world in which companies and individuals are made and
destroyed on social networks; that’s how powerful these platforms are. User expecta‐
tions are fueling complexity in applications and infrastructure, so much so that over
50% of mobile users abandon sites that take more than three seconds to load. Most
web pages take a lot longer than that to load, creating a significant gap between con‐
sumers’ expectations and most businesses’ mobile capabilities. If you’re an
175
organization that services customers and you know of this stat, then you must ensure
that you have a system that’s constantly monitoring your website load times and alert‐
ing you when numbers are outside acceptable bounds. You want to make sure that
your team can resolve issues before they become bigger problems.
So what is monitoring? In simple terms, monitoring is a comprehensive operations,
policies and performance management framework. The aim is to detect and alert
about possible errors of a program or a system in a timely manner and deliver value
to the business. Organizations use monitoring systems to monitor devices, infrastruc‐
ture, applications, services, policies, and even business processes. Because monitoring
applies to many areas of the business beyond what we can cover, in this chapter we
will primarily focus on monitoring as it relates to governance.
Monitoring governance involves capturing and measuring the value generated from
data governance initiatives, compliance, and exceptions to defined policies and pro‐
cedures—and finally, enabling transparency and auditability into datasets across their
life cycle.
What’s interesting about monitoring is that when everything is working well and
there are no issues, efforts usually go unnoticed. When issues arise and things go
wrong—for example, with data quality or compliance exceptions—then governance is
one of the first areas that gets blamed because these areas fall right within governance
areas that should be constantly monitored. (We will do a deeper dive into each of
these areas later in the chapter.) Because of this, it’s important for you to define met‐
rics that allow you to demonstrate to the business that your efforts and investments in
data governance are benefiting the business in reducing costs, increasing revenue,
and providing business value. You need to implement metrics that you will track
accordingly, enabling you to understand and showcase the improvements you are
making to the bottom line. In later sections of this chapter, we will provide you with a
list of some of the key metrics you can track to show the efficacy of your governance
program.
Why Perform Monitoring?
Monitoring allows you to review and assess performance for your data assets, intro‐
duce policy changes within the organization, and learn from what’s working and
what’s not, with the ultimate goal of creating value for the business. If your organiza‐
tion is used to hearing about incidents via its customer base and is spending too
much time and money on manual support, then a monitoring system is vital. Moni‐
toring serves many different functions, and for the majority of the use cases, a moni‐
toring system will help you with alerting, accounting, auditing, and compliance. Let’s
do a deep dive into these core areas:
176
|
Chapter 8: Monitoring
Alerting
In its simplest terms, an alert warns someone or something of a danger, threat, or
problem—typically with the intention of avoiding it or dealing with it. An alert
system can help you prevent incidents, and when incidents happen, they are
detected earlier and faster. A governance monitoring system that is specifically
designed to monitor data quality can alert you when data quality thresholds fall
outside the allowable limits, allowing you to avoid falls in service or to minimize
the amount of time needed to resolve the issue.
Accounting
An account refers to a report or description of an event or experience. In this
core area of monitoring, you want to get an in-depth analysis of your applica‐
tions, infrastructure, and policies. This enables you to create more appropriate
strategies, set realistic goals, and understand where improvements are needed,
allowing you to discover the effectiveness and value generated from data gover‐
nance and stewardship efforts.
Auditing
Auditing refers to a systematic review or assessment of something to ensure that
it is performing as designed. Audits also enable transparency into data assets and
their life cycle. Auditing allows you to understand the ins and outs of your busi‐
ness so that you can make improvements to processes and internal controls, with
the aim of reducing organizational risk and preventing unexpected costs from
external auditors.
Compliance
Regulatory compliance refers to the efforts required to help you meet relevant
policies, laws, standards, and regulations. When your systems, thresholds, or pol‐
icies are outside the defined rules, resulting in out-of-compliance processes,
monitoring can help you stay in compliance; your business can be alerted as soon
as these exceptions are detected, giving you the opportunity to resolve the issues
in order to stay compliant.
Although this is not the first time we’re covering this subject, we
want to reiterate just how important alerting is to your governance
program. In countless interviews, and during our research, the lack
of sufficient alerting has been mentioned as a top pain point among
data scientists, data engineers, and the like. While there are many
ingredients we’ve discussed that are key to a successful governance
strategy, proper alerting is something that often gets overlooked.
Improvements in this area not only aid in prevention and early
detection of incidents but also help to streamline tasks and the time
spent on those tasks by your employees.
Why Perform Monitoring?
|
177
Monitoring use cases will tend to fall within the four buckets we’ve just highlighted,
and in most cases they will overlap. For example, as you monitor compliance in your
organization, this can also help you when it comes to audits and proving that your
organization is doing the right things and performing as designed.
Monitoring is often done using a monitoring system that can be built in-house by
purchasing a system and configuring it to your other internal systems. It can be out‐
sourced to other vendors as a managed service for ease of setup and expense, or if
you’re using cloud solutions, it can be embedded within your organization’s work‐
flows. Open source tools also provide some monitoring solutions to be considered.
We will delve more deeply into monitoring systems later in this chapter; for now, let’s
focus on what areas of governance you should monitor, why, and how.
Why Alerting Is a Critical Monitoring Function
This is a cautionary tale about a small website management services company that
served 50 major hospitals around the country. The story sounds a little too familiar,
and honestly, this continues to happen to many organizations that store customer
information. For the company, things were going so well that it decided to introduce a
new service that allowed patients and partners to do online bill payment.
One day during a maintenance procedure, an employee accidentally turned off the
company’s firewall, and all of the patient data stored on its online billing service sys‐
tem (for at least five hospitals and approximately 100,000 patients) was left exposed to
the world.1 And because the hospitals reported the breaches as separate incidents, the
nature of the error was not immediately apparent.
Unfortunately for this small company, there was no way to recover from such an inci‐
dent. It simply closed its doors and shut down its website. There’s so much to learn
from this experience, because if customer data had been properly governed and
stored—that is, if the data had been encrypted and the right access controls had been
enforced—it would have been easier to catch the incident right when it happened and
alert the hospitals as needed. In addition, this reinforces the importance of monitor‐
ing and, more specifically, of alerting. It’s scary to think that an employee turned off a
firewall system without any system alerts being triggered or the right folks being
alerted.
1 Tim Wilson, “A Cautionary Tale”, Dark Reading, August 17, 2007.
178
|
Chapter 8: Monitoring
What Should You Monitor?
There are many areas of an organization that are monitored: operating systems and
hardware, network and connectivity, servers, processes, governance, and more. In
order to stay close to the core of this book and chapter, we will now do a deeper dive
into monitoring as it relates to data governance. A lot of the concepts we will explore
next have already been covered in previous chapters, and so we will focus on which
parts need to be monitored and how to go about it. We will keep referring back to
what you’ve learned in order to solidify these concepts across the key areas.
Data Quality Monitoring
In Chapters 1 and 2, we introduced the concept of data quality, and we went into that
in more detail in Chapter 5. Data quality allows the organization to trust the data and
the results. High-quality data means that the organization can rely on that data for
further calculations/inclusions with other datasets. Because of how important data
quality is, it should be monitored proactively, and compliance exceptions should be
identified and flagged in real time. This will allow the organization to move quickly to
identify and mitigate critical issues that can cause process breakdowns.
There are critical attributes of data quality that should be tracked and measured,
including completeness, accuracy, duplication, and conformity. These are outlined in
detail in Table 8-1 and should map to specific business requirements you have set
forth for your governance initiative. Monitoring can help you create controls for vali‐
dation and can provide alerts as needed when these are outside the defined thresh‐
olds. The attributes in Table 8-1 are common problem areas within the data
management space that make it difficult to trust data for analysis.
Table 8-1. Data quality attributes
Attribute
Description
Completeness This identifies what data is missing and/or not usable.
Accuracy
This identifies the correctness and consistency of the data and whether the correct values are stored for an
object.
Duplication
This identifies which data is repeated. Duplicate records make it difficult to know the right data to use for
analysis.
Conformity
This identifies which data is stored in a nonstandard format that will not allow analysis.
Process and tools for monitoring data quality
A data quality monitoring system routinely monitors and maintains data quality
standards across a data life cycle and ensures they are met. It involves creating con‐
trols for validation, enabling quality monitoring and reporting, accessing the level of
What Should You Monitor?
|
179
incident severity,
recommendations.
enabling
root
cause
analysis,
and
providing
remedy
When setting up a data quality monitoring process, some of the things you need to
consider are:
Establishing a baseline
Establish a baseline of the current state of data quality. This will help you identify
where quality is failing and help you determine what is “good” quality and what
those targets are. These targets must be tied to the business objectives you have
set forth for your governance initiatives. Comparing results over time is essential
to proactive management of ongoing data quality improvement and governance.
Quality signals
These are usually monitored over a period of time or by the source of the data.
The monitoring system will be looking to verify data fields for completeness,
accuracy, duplicates, conformity, statistical anomalies, and more. When data
quality falls below a specified threshold, an alert would be triggered with more
information about the quality issue observed. These quality signal rules are usu‐
ally set forth by the data governance committee, which ensures compliance with
data policies, rules, and standards. More details around this are outlined in Chap‐
ter 4. These policy guidelines and procedures ensure that the organization has
the data program that was envisioned by the organization.
In order to get started with monitoring data quality, determine a set of baseline met‐
rics for levels of data quality and use this to help you build a business case to justify
the investment and, over time, help you make improvements to the governance
program.
Data Lineage Monitoring
We introduced the concept of data lineage in Chapter 2 and talked about why track‐
ing lineage is important. The natural life cycle of data is that it is generated/created by
multiple different sources and then undergoes various transformations to support
organizational insights. There is a lot of valuable context generated from the source of
the data and all along the way that is crucial to track. This is what data lineage is all
about. Monitoring lineage is important to ensure data integrity, quality, usability, and
the security of the resulting analysis and dashboards.
Let’s be honest: tracking and monitoring lineage is no simple task; for many organiza‐
tions, their data life cycle can be quite complex, as data flows from different sources,
from files to databases, reports, and dashboards, while going through different trans‐
formation processes. Lineage can help you track why a certain dashboard has differ‐
ent results than expected and can help you see the movement of sensitive data classes
across the organization.
180
|
Chapter 8: Monitoring
When looking to track and monitor lineage, it’s important to understand certain key
areas that you will come across. Table 8-2 has some of these attributes. These are not
all encompassing but are just some of the more important ones for you to know.
Table 8-2. Data lineage attributes
Attribute
Data
transformations
Technical metadata
Data quality test
results
Reference data
values
Actor
Description
These are the various changes and hops (aggregates, additionals, removals, functions, and more) as
data moves along the data life cycle. Monitoring helps you understand details of the data points and
their historical behavior, as data gets transformed along the way.
Metadata is important for understanding more about the data elements. Enabling automatic tagging
of data based on its source can help provide more understanding of the data asset.
These represent data quality measurements that are tracked at specific points of the data life cycle, in
order to take action as/if needed.
Reference data values can be used to understand backward data lineage and transformation from that
specific data point, and/or the intermediate transformation following that point with forward data
lineage. Understanding reference data values can be useful in performing root cause analysis.
An actor is an entity that transforms data. It may be a MapReduce job or an entire data pipeline. Actors
can sometimes be black boxes, and the inputs and outputs of an actor are tapped to capture lineage in
the form of associations.
Process and tools for monitoring data lineage
What makes monitoring lineage so complicated is that it needs to be captured at mul‐
tiple levels and granularities—this is tedious and time consuming because of all the
intricacies and dependencies. As mentioned earlier in this chapter, monitoring serves
different purposes; for lineage it can be used to alert, audit, and comply with a set of
defined rules and policies. As described in Table 8-2, one of the things you could
monitor is the behavior of the actors; when their resulting transformed outputs are
incorrect, an alert function can be set up that the inputs are investigated, and the
actors are augmented and corrected to behave as expected or are removed from the
process flow of the data.
Monitoring lineage can also provide a lineage audit trail, which can be used to deter‐
mine the who, what, where, and when of a successful or attempted data breach in
order to understand which areas of the business were or might have been affected by
the breach. In addition, the important details tracked by data lineage are the best way
to provide regulatory compliance and improve risk management for businesses.
Lineage is about providing a record of where data came from, how it was used, who
viewed it, and whether it was sent, copied, transformed, or received and it is also
about ensuring this information is available. You will need to identify the best way to
do this for your organization, depending on the use cases and needs of the business.
Identify the data elements, track the listed elements back to their origin, create a
repository that labels the sources and their elements, and finally, build visual maps for
each system and a master map for the whole system.
What Should You Monitor?
|
181
Compliance Monitoring
Compliance has been covered in great detail in several of the previous chapters.
Understanding state and federal regulations, industry standards, and governance pol‐
icies and staying up to date on any changes ensures that compliance monitoring is
effective.
The changing nature of laws and regulations can make monitoring compliance diffi‐
cult and time consuming. Noncompliance is not an option because it often results in
considerable fines—sometimes more than twice the cost of maintaining or meeting
compliance requirements. In a study from the Ponemon Institute, the average cost for
organizations that experience noncompliance problems is $14.82 million. That
includes fines, forced compliance costs, lack of trust from customers, and lost busi‐
ness.
Monitoring compliance means having an internal legal representative (attorney),
though at times this task falls on the privacy tsar or someone else in security, who
must continually keep up with laws and regulations and how they impact your busi‐
ness. It also requires that changes are made according to the new information gath‐
ered in order to stay in compliance. In addition, to stay in compliance requires
auditing and tracking access to data and resources within the organization. All of this
is done to ensure your business is compliant in case of an audit by the government.
Process and tools for monitoring compliance
To be successful in monitoring regulatory compliance, you need to evaluate which
regulations apply to your data governance efforts and what compliance looks like
with these regulations. Once this is done, do an audit to understand your current
governance structure with regard to the relevant regulations. Consider this a bench‐
mark exercise to understand your current standing, future requirements, and how
you can build a plan to get to the end state, and then continuously monitor that end
state to ensure compliance.
Once you’ve completed the audit, you should jump to creating a monitoring plan;
here you should address all risks identified in the audit stage and prioritize those that
are the greatest threat to the organization. Next, decide how you are going to imple‐
ment the monitoring program, including roles and responsibilities.
The resulting output will be dependent on the level and frequency of regulatory
changes and updates from the relevant regulatory boards. Make sure you have a way
to inform the regulatory agencies of failed audits and of how you’re looking to miti‐
gate them.
Here are some ways you can be proactive about compliance:
182
|
Chapter 8: Monitoring
• Keep on top of regulatory changes and make sure that you’re checking for up
dated standards and regulations. Of course, this is easier said than done.
• Be transparent so your employees understand the importance of compliance and
the regulations they are complying with. Offer training sessions to explain the
regulations and their significance.
• Build a culture of compliance within the organization—and yes, there must be
someone who has the task of staying up to date on regulatory requirements that
the company may (or may not) need to be in compliance with. Most big compa‐
nies have large compliance teams; even if you’re a small organization; however,
it’s still important to designate someone to handle compliance, including moni‐
toring, checking for updates in regulations and standards, and more.
• Cultivate a strong relationship between your compliance person/team and the
legal department so that when incidents occur, those teams are in lockstep with
each other and are already used to working together.
Monitoring compliance can be a manual process, with checklists and all, but that can
also make things even more complicated. There are automated compliance tools that
your organization can look at that provide compliance in real time, giving you con‐
tinuous assurance and minimizing the chance that human error may lead to a gap in
compliance.
Program Performance Monitoring
Monitoring and managing the performance of a governance program is integral to
demonstrating program success to business leadership. This type of monitoring
allows you to track progress against the program’s aims and objectives, ensuring the
governance program delivers the right outcomes for the organization, accounting for
efficient and effective use of funding, and identifying improvement opportunities to
continue creating impact for the business.
In order to monitor program performance, some of the items you might want to
measure include:
• Number of lines of business, functional areas, and project teams that have com‐
mitted resources and sponsorship
• Status of all issues that come into the governance function, how they were han‐
dled, and what the resulting impact was
• Level of engagement, participation, and influence the governance program is
having across the organization, which will help people understand the value of
the governance program
What Should You Monitor?
|
183
• Value-added interactions, including training and project support, and what the
impact is to the business
• Business value ROI from data governance investments, including reducing pen‐
alties by ensuring regulatory compliance, reducing enterprise risk (e.g., contrac‐
tual, legal, financial, brand), improving operational efficiencies, increasing topline revenue growth, and optimizing customer experience and satisfaction
Process and tools for monitoring program performance
Program performance monitoring needs to be continuous and ongoing, helping you
identify areas that are not performing to expectations and determining what types of
program adjustments are needed. Most performance management frameworks con‐
sist of the sets of activities outlined in Figure 8-1.2
Figure 8-1. Performance management framework
Let’s look more closely at each of these areas:
Alignment with existing governance frameworks
This ensures that your program performance is aligned with the established gov‐
ernance framework. In Chapter 4, we did a deep dive into governance frame‐
works, groups, and what is needed for effective governance.
Developing performance indicators
Now that you are aligned with existing governance frameworks, develop key per‐
formance indicators (KPIs) for your governance program. These should be well
defined, relevant, and informative.
Reporting progress and performance
Documenting the governance performance objectives and how they’re being met
and sharing this information with leadership will ensure that people see the value
of the program. Providing reports becomes an essential way for people to con‐
sume this information.
2 Grosvenor Performance Group, “How Is Your Program Going…Really? Performance Monitoring”, May 15,
2018.
184
|
Chapter 8: Monitoring
Taking action based on performance results
It’s important to identify ways to ensure that the performance results are used to
inform decision making in your organization; otherwise, what is the point?
As you can see in Figure 8-1, this is an ongoing and iterative process, with one area
feeding and informing the other.
Security Monitoring
Cyberattacks are becoming bigger than ever before, with new threats from state
actors and increasingly sophisticated methods. The damage related to cybercrime is
projected to hit $6 trillion annually by 2021, according to Cybersecurity Ventures.3 In
addition, many countries are now taking consumer data privacy and protection more
seriously by introducing new legislation to hold businesses accountable.
Attacks cost more than money. They can affect a business’s brand and shareholder
reputation. In the recent Equifax data breach, in which over 140 million records were
exposed, the company most likely incurred a cost of more than $32 billion to resolve
the issue.4 This was reflected in their stock price, which fell more than 30% after the
breach. Perhaps even worse, adjacent firms in Equifax’s industry who did not get
breached felt a 9% stock drop, likely due to loss of confidence in security measures.5
So even if you’re doing everything right, you can still be impacted by a breach.
That’s why security monitoring is so important. It is the process of collecting and ana‐
lyzing information to detect suspicious behavior, or unauthorized system changes on
the network in order to take action on alerts as needed. Most companies are routinely
exposed to security threats of varying severity; the causes of security breaches include
hackers and malware, careless employees, and vulnerable devices and operating sys‐
tems. Security threats are part of the normal course of conducting business; therefore
it’s important to be prepared and act on threats and breaches before they cause dam‐
age and disruption.
There are many areas of the business where you can monitor security. Table 8-3 high‐
lights some of those areas.
3 “29 Must-Know Cybersecurity Statistics for 2020”, Cyber Observer, December 27, 2019.
4 “5 Reasons Why You Need 24x7 Cyber Security Monitoring”, Cipher (blog), May 15, 2018.
5 Paul R. La Monica, “After Equifax Apologizes, Stock Falls Another 15%”, CNNMoney, September 13, 2017.
What Should You Monitor?
|
185
Table 8-3. Security monitoring items, showing some areas of the organization that you can
monitor security on
Item
Security alerts and
incidents
Network events
Server logs
Application events
Server patch
compliance
Endpoint events
Identity access
management
Data loss
Description
These are any alerts or incidents generated from an IT environment. They could be data exfiltration or
unusual port activity, acceptable use policy violations, or privileged user-activity violations.
This involves the ability to monitor network activity and receive alarms or reports of the occurrence of
selected events, including device statuses and their IP addresses, new device alerts, and network
status.
This involves monitoring and detecting server activities continuously, examining alerts before a server
mishap occurs, recording server logs and reports for easy tracking of errors, performing log analysis,
and monitoring server performance and capacity.
This involves monitoring events surrounding your software and applications in their stack, ensuring
they are accessible and performing smoothly.
This involves installing and patching all the servers in your IT environment and staying compliant. This
helps mitigate vulnerabilities, server downtimes and crashes, and slowing down.
This is a list of all events that can be emitted by an instance of an application, a process, or an event.
This involves defining and managing the roles and access privileges of individual network users and
maintaining, modifying, and monitoring this access throughout each user’s access life cycle.
This involves detecting potential data breaches/data exfiltration transmissions and employing
monitoring techniques and prevention when data is in use, in motion, and at rest.
Process and tools for monitoring security
An effective process of monitoring security is continuous security monitoring, pro‐
viding real-time visibility into an organization’s security posture and constantly look‐
ing for cyber threats, security misconfigurations, and other vulnerabilities. This
allows an organization to stay a step ahead of cyber threats, reducing the time it takes
to respond to attacks while complying with industry and regulatory requirements.
Cyber security can be conducted at the network level or the endpoint level. With net‐
work security monitoring, tools aggregate and analyze security logs from different
sources to detect any failures. Endpoint security technologies, on the other hand, pro‐
vide security visibility at the host level, allowing for threats to be detected earlier in
the process flow.
Security monitoring is an area with many players: from companies that offer solu‐
tions you can implement within your organization to full-fledged companies that you
can outsource the entire service to. The solution you choose to go with depends on
your business, the size of your in-house team, your budget, the technologies at your
disposal, and the level of security-monitoring sophistication that you’re looking for.
Both options have pros and cons that need to be weighed before you can make an
effective decision for your business.
186
|
Chapter 8: Monitoring
You should now have an understanding of what governance items to monitor, the
how, and the why. The next section will look more closely at monitoring systems,
their features, and which criteria to monitor.
What Is a Monitoring System?
Monitoring systems are the core set of tools, technologies, and processes used to ana‐
lyze operations and performance in order to alert, account, audit, and maintain the
compliance of organizational programs and resources. A robust monitoring system is
paramount to the success of a program and needs to be optimal with regard to what
the business needs.
Monitoring can be done in-house by purchasing a system and configuring it to your
other internal systems; it can be outsourced as a managed service to other vendors
because of an internal lack of expertise and expense; or if you’re using cloud solu‐
tions, it can be embedded within your organization’s workflows. Open source tools
also provide some monitoring solutions that can be considered as well. Whatever
option you choose to go with, here are common features of a good monitoring
system.
Analysis in Real Time
Real time is real money. In a world in which things change so fast and people need
information at their fingertips, you must have a monitoring system that does analysis
in real time. A good system should offer continuous monitoring with minimal delays,
allowing you to make changes and improvements on the fly.
System Alerts
A monitoring system needs to have the ability to signal when something is happening
so that it can be actioned. A system that allows multiple people to be informed with
the right information will go a long way toward ensuring that issues are addressed as
quickly as possible. A system should allow configuration for multiple events and have
the ability to set different sets of actions depending on the alert. The alert should con‐
tain information about what is wrong and where to find additional information.
Notifications
A good monitoring system needs to have a robust, built-in notification system. We’re
no longer in the age of pagers, so your system needs to be able to send SMS, email,
chat, and more to ensure that the message gets to the right folks. And once the mes‐
sage is received, the right people need to be able to communicate back to the team
that the alert has been received and that the issue is being investigated—and when
issues are resolved, communicate that the system or process is back to normal
What Is a Monitoring System?
|
187
operations. Notifications could kick off additional processes automatically, where cer‐
tain systems take an action.
Reporting/Analytics
Monitoring systems are big data collection and aggregation units given all the alerts
and trigger events collected over a period of time. Your monitoring system needs to
allow for robust reporting that will allow you to present the data to clients or different
departments in the organization. Reporting allows you to identify trends, correlate
patterns, and even predict future events.
Graphic Visualization
Collecting data should be augmented with the ability to analyze and visualize a situa‐
tion. Dashboards play a critical role in ensuring that everyone has a place to look to
see how things are going, and even to observe trends over time. A visual representa‐
tion of what’s happening is easier for people to understand and absorb and is one of
the top customer requests in a data governance solution. A good monitoring system
should have friendly and easy-to-understand graphs that allow the organization to
make decisions.
Customization
It’s no surprise that different organizations have different business requirements. You
need to have the ability to be able to customize your monitoring system by function,
user type, permissions, and more, allowing you to have the right alerts triggered, and
to have them actioned by the right folks.
Monitoring systems need to run independently from production services, and they
should not put a burden on the systems they are tracking. This simply allows the
monitoring system to continue running in the case of a production outage or other
failure event. Otherwise, a failure could take down a monitoring system (when the
production environment is down), and that defeats the purpose of its existence. In
addition, as you would with any system, ensure you have a failover for your monitor‐
ing system in case it is the one that’s affected by a blackout or failure. Figure 8-2 high‐
lights a simple monitoring system, given some of the features we’ve just highlighted.
188
|
Chapter 8: Monitoring
Figure 8-2. Example monitoring system
As you can imagine, monitoring systems are becoming even more sophisticated with
the introduction of machine learning capabilities, so even though this list of features
is robust, it’s not by any measure the be-all and end-all. Use this list as a starting point
to select or build the right monitoring system for your organization, and, depending
on your use case and company needs, augment your system as needed.
Monitoring Criteria
Now that you have selected a monitoring system, what are the types of criteria that
you can set up to monitor? Monitoring systems collect data in two distinct ways: pas‐
sive systems, where the tools observe data created by the application and system under
normal conditions (i.e., log files, output messages, etc.); and active systems, which are
more proactive, use agents and other tools to capture data through a monitoring
module, and are often integrated within production systems.
Here’s some common criteria you can follow as you set up your monitoring system:
Basic details
Identify basic details for the items you’re looking to monitor. Capture a list of
rules, their purpose, and a distinct name for each one. If your system has prede‐
fined alerts and queries, select the ones you want to monitor and label them
accordingly so they can be easily identified.
Monitoring Criteria
|
189
Notification condition
Program your notification condition. Once the query/criteria you set is identified
and the result is evaluated against configurable thresholds, an alert should be set
off if there’s a violation on the criteria set forth. Users should be identified by
email or SMS, and the information should be made available on the monitoring
dashboard. In addition, it’s not enough to simply send alerts; you need to ensure
that someone acknowledges that they’ve received the alert and that they are
working on it. Also, the alert needs to go to someone on call, and if that person,
for whatever reason, is not available, it needs to be routed to the right person who
can respond to the alert. If the alert is not responded to, the system should send
another notification within a specified amount of time as long as the metric con‐
tinues to remain outside the threshold.
Monitoring times
In this instance, specify the frequency of the running of a certain validation
(daily/weekly/monthly), and then how long and how frequently it should occur
within the day (business hours within the day and frequency within the day).
This is more of a gut check system that looks to make sure processes and systems
are operating correctly; it is more useful for passive systems in which things are
monitored to ensure ongoing operations.
Given the complexity of all these items and the level of sophistication needed to
maintain them, monitoring systems are what make all this possible. The next section
offers some reminders and touches on other things you should keep in mind.
Important Reminders for Monitoring
From reading this chapter, you will have discerned some recurring themes when it
comes to monitoring. Here are some considerations to keep in mind:
Getting started with a monitoring system
Just like any software development process, monitoring can be done in-house by
purchasing a system and configuring it to your systems; it can be outsourced as a
managed service to other vendors because of an internal lack of expertise and
expense,; or if you’re using cloud solutions, it can be embedded within your
organization’s workflows. Open source tools also provide some monitoring solu‐
tions that can be considered as well.
Real-time improves decision making
For the majority of the governance items outlined in the previous section, a con‐
tinuous, real-time monitoring system is paramount to improving decisionmaking and staying on top of compliance. In addition, having an alert system
that is robust and allows people to take action as needed will ensure that things
are remedied within the outlined SLAs.
190
|
Chapter 8: Monitoring
Data culture is key to success
Employee training needs to be embedded into the fiber of your business—your
data culture, as discussed in Chapter 9. Many governance concerns are under‐
taken by employees who might not be aware that they’re doing something that
will compromise the business. Find ways to make education personal, because
people are most likely to apply education when they see it as important to their
lives.
Summary
This chapter was designed to give you a foundation on monitoring and how to think
about implementing it for your organization. Monitoring is vital to understanding
how your governance implementations are performing on both a day-to-day and a
longer-term basis. Monitoring is where the rubber meets the road, allowing you to
ask for additional resources, course-correct as needed, learn from the wins and fail‐
ures, and really showcase the impact of your governance program.
Take an audit of your current governance and monitoring initiatives and augment
them as needed, because doing this can only reap more benefits for your
organization.
Summary
|
191
CHAPTER 9
Building a Culture of Data Privacy
and Security
In this book we have covered a lot: considerations in data governance, the people and
processes involved, the data life cycle, tools, and beyond. These are all pieces of a puz‐
zle that need to come together in the end for data governance to be successful.
As alluded to earlier, data governance is not just about the products, tools, and people
who carry out the process—there is also the need to build a data culture. The careful
creation and implementation of a data culture—especially one focused on privacy
and security—not only aids in the establishment of a successful data governance pro‐
gram but also ensures that the program is maintained over time.
As we discussed in Chapter 3, the people and processes around data governance—in
conjunction with data governance tools—are essential components of a data gover‐
nance program. In this chapter we will go one step further than thatby including the
data culture (the culture within a company around data) as a key final component in
creating a highly successful data governance program.
Data Culture: What It Is and Why It’s Important
The data culture is the set of values, goals, attitudes, and practices around data, data
collection, and data handling within a company or organization. While many compa‐
nies or organizations spend quite a bit of time vetting data governance tools and cre‐
ating processes, they often fail to set up a culture within the company around data.
This culture defines and influences things like:
193
• How data is thought about within the company/organization (Is it an asset?
Needed to make decisions? The most important part of the company? Just some‐
thing to be managed?)
• How data should be collected and handled
• Who should be handling data and when
• Who is responsible for data during its life cycle
• How much money and/or resources will be committed to serving the company/
organization’s data goals
While this is certainly not an exhaustive list, it begins to show you the vast amount of
considerations that go into the definition of a data culture.
Having a set and defined data culture is important for a multitude of reasons, but the
most important is that it sets the stage and serves as the glue that holds everything
else together within a data governance program.
We will go into more detail in this chapter about what the “North Star” of a data cul‐
ture should look like and about what considerations you should be making when
designing your own.
Starting at the Top—Benefits of Data Governance to the
Business
A key aspect of building a successful data culture is getting buy-in from the top down,
which generally requires that decision makers within a company see how a data gov‐
ernance program will work and agree on why implementation of a data culture bene‐
fits the company’s bottom line. An efficient data culture aids in ensuring reliable,
high-quality data that not only produces better analytics but also reduces compliance
violations and penalties. All of this will result in better business performance, as high‐
lighted in various studies including a 2018 McKinsey report that found that “break‐
away companies” are twice as likely to claim they had a strong data governance
strategy.1 In many of our interviews with companies, one of the most common com‐
plaints we hear about initiating and implementing a data governance program is get‐
ting buy-in from those that have the power to fund data governance initiatives, as
well as to support and enforce a data culture.
1 Peter Bisson, Bryce Hall, Brian McCarthy, and Khaled Rifai, “Breaking Away: The Secrets to Scaling Analyt‐
ics”, McKinsey & Company, May 22, 2018.
194
| Chapter 9: Building a Culture of Data Privacy and Security
Often the investment in a data governance program and building a data culture (both
in terms of the purchasing of tools and infrastructure and in terms of headcount) is
seen as a cost with no ROI other than hoped-for assurance that a company won’t be
hit with fines for being out of compliance with a regulation. A well-thought-out and
well-executed governance program can, however, provide the cost savings of proper
data handling, and it also has the ability to increase the value of already existing assets
(“old data” that may be sitting around).
We have found and will discuss several areas that can be quite persuasive in helping
decision makers to see the value and importance of building a data culture.
Analytics and the Bottom Line
We’ve discussed in depth the implications of governance and how knowing what data
there is, and where it is, not only aids in locking down and handling sensitive data
but also helps analysts to run better, more actionable analytics. Better analytics—ana‐
lytics that are based on higher-quality data or come from a synthesis of data from
multiple sources, for example—enable us to make better data-driven decisions. All
companies are driven by a desire to increase their profitability, whether by increasing
revenue and/or by decreasing waste or expenditure. The entire push to data-driven
decision making has at its heart the hope and promise of driving revenue. When deci‐
sion makers are able to see that a data governance program and data culture generate
better analytics, which in turn positively affects the bottom line, they are not only a
bit more likely to be passively “bought in,” but they may also rise to be champions
and drivers of the data culture.
Company Persona and Perception
While not directly related to the bottom line, there is much to be said for the public
perception of a company and how it handles data.
In the last five years, there have been several companies that have fallen under great
scrutiny regarding the data they collect and how that data is used. The public’s per‐
ception that these companies have used data unethically carries a host of cascading
negative repercussions—ranging from decreased employee morale to financial effects
such as dropped sponsors and/or losing customers to a competitor—and these do
indeed affect a company’s bottom line.
Starting at the Top—Benefits of Data Governance to the Business
|
195
Top-Down Buy-In Success Story
One company we’ve interviewed has had a particularly successful execution of a data
governance program, in large part due to the buy-in it received from the top of the
organization. This company works in research and in healthcare and is looking to lev‐
erage analytics across data collected on each side of the business. Currently its data
resides in separate storage for each line of business. To marry this data together into a
central repository, it recognized the need for a comprehensive data governance
strategy.
We hope we’ve impressed upon you in this book that all companies should have a data
governance strategy and program, but of course there are certain business types that,
due to their high level of regulation, need to be even more conscious of their imple‐
mentation of governance. Healthcare is one of those. This company knew that a com‐
prehensive governance program would need to be broad and be executed at many
different levels within the organization. The company deals almost exclusively in sen‐
sitive data, and for this data to be moved, manipulated, joined with other datasets,
and so on, the governance on it had to be top-notch.
To achieve the level of buy-in that would be required, the company set out to create a
charter that outlined exactly what the governance program within the company
should look like: the framework/philosophy that should be followed, tools that would
be needed, the headcount needed to execute these tools, and notably, the establish‐
ment and ongoing reinforcement of a data culture. Is this exhaustive? Lofty? Idealis‐
tic? Perhaps. But through this charter, the company was able to gain that buy-in from
the top, and it is now executing one of the most well-thought-out and structured gov‐
ernance programs we’ve seen (complete with headcount for governance-only related
tasks—a true rarity).
We include this example not to imply that every single one of these “boxes” must be
ticked or that a smaller-scale or fledgling governance program is not worth it. On the
contrary: we hope that this example (while extreme) serves to show you just how
much documenting a well-thought-out strategy, including how this strategy will be
embedded into the culture of the company, can help you in getting not only initial
buy-in, but also the level of buy-in that will maintain and support your governance
program in the long run.
Intention, Training, and Communications
Perhaps one of the most important aspects of building a data culture is the internal
data literacy, communications, and training. In the overview of this chapter, we men‐
tioned that a key to successful governance and data culture is not just the establish‐
ment of a program but maintenance over time to ensure its longevity. Integral to this
196
|
Chapter 9: Building a Culture of Data Privacy and Security
end is intention, data literacy (a deep understanding of and ability to derive meaning‐
ful information from all kinds of data), training, and communications.
A Data Culture Needs to Be Intentional
Just as we discussed that a successful governance program needs to have a set process,
the building and maintenance of a data culture is very much the same.
What’s important
When creating and establishing a data culture, a company first needs to decide what’s
important to it—what its tenets are. For example, a company that deals with a lot of
sensitive data (like a healthcare company) may decide that the proper treatment and
handling of PII is a primary tenet, whereas a small gaming app company may decide
that ensuring data quality is its primary focus. Alternatively, a company may decide
that many tenets are important to it, which is fine—the key aspect here is that these
tenets need to be clearly defined and collectively agreed upon as the rest of the estab‐
lishment and maintenance of a data culture stems from this step.
Aside from internal tenets, there also exist tenets that are nonnegotiable in nature,
such as legal requirements and compliance standards that all or most companies need
to integrate as tenets, no matter what.
It’s also worth noting here that in addition to tenets and requirements/compliance
standards, we would argue that an important and perhaps even crucial tenet is that of
caring. This may seem like a strange tenet to have as part of a data governance pro‐
gram, but it is an essential component of a data culture. For a governance program to
function at its highest level, there must be an intrinsic desire on the part of the com‐
pany and its employees to do the right thing. The protection of and respect for data
has to be an integral part of the fabric of the company and its data-handling ethos. It
must be something that is touted and supported from the top so that it cascades down
to the rest of the company. Without this concern, the company is one that merely
deals with data, and perhaps has a data governance program but has no data culture.
While no company wants to rely solely on its employees to do the right thing, foster‐
ing a data culture helps to fill in the gaps for when things invariably go sideways.
Training: Who Needs to Know What
A successful implementation of a data culture not only needs to have its tenets well
defined; it also needs to identify who will be carrying these out, how they will do this,
and whether or not they have the knowledge and skills needed for proper execution.
Intention, Training, and Communications
|
197
The “who,” the “how,” and the “knowledge”
Too often we have seen that one or more of these three areas is glossed over or taken
for granted. As when we discussed the different “hats” involved in data governance,
there is a high likelihood that several of the folks carrying out these tasks have: littleto-no technical knowledge; some may be doing other tasks as part of their role, leav‐
ing little time to dedicate to data governance; and there are often breakdowns around
who is responsible for what tasks. Each of these components is important and should
not be overlooked. Each component needs to be thoroughly considered when plan‐
ning and implementing the data culture aspect of a data governance program.
Part of the data governance and data culture execution strategy includes determining
who will do what. Not only is this important in terms of defining roles and responsi‐
bilities, but it’s also important that the people who will be fulfilling these duties have
the skills and the knowledge to perform their tasks. It is simply not enough to enlist a
person to a task, give them a tool, and hope for the best. A plan for how knowledge
and skills will be not only acquired but also maintained over time is critical.
A solid plan for what training will be required and how it will be disseminated is
essential. Again, this is a place where we’ve seen companies falter—they may have
decided who does what and even provided some initial training to get people up to
speed, but they forget that as technology and the data collected grow and change, new
skills and knowledge may be necessary. Training is often not a “one-and-done” event
—it should be seen as ongoing.
In devising your own training strategy, you should think in terms of what’s initially
needed to teach your staff and what should continue to be reinforced and/or intro‐
duced in subsequent training. For example, we have seen companies have success
with setting up events such as “Privacy Week” or “Data Week,” in which essential
training about proper data handling and governance considerations are reviewed and
new considerations and/or regulations are introduced. What makes these “events” so
much more successful than simply providing required virtual click-through training
is that you can center your event around a particular topic of recent importance (per‐
haps an internal issue that occurred, a new regulation that’s been launched, or even a
highly publicized issue from an external company). This event structure gives you
some freedom and flexibility around how to go about your training, depending on
what’s most important to either reinforce or disseminate to your staff.
An example of what this type of event could look like is shown in Table 9-1.
198
|
Chapter 9: Building a Culture of Data Privacy and Security
Table 9-1. Sample schedule
Monday
Tuesday
“Basics of Proper Data Handling and Governance” (one hour)
“How to Use Our Governance Tools 101”
or “Advanced ‘How to’ on Our Governance Tools. Did You Know You Can Do…?!” (one hour)
Wednesday “Governance and Ethics: How and Why Governance Is Everyone’s Responsibility” (one hour)
Thursday
“How Do I Handle That? A Guide On What to Do When You Encounter a Governance Concern” (one hour)
Friday
Guest speaker on an aspect of governance and/or data culture you find particularly important to your
organization (one hour)
Of course you will have to evaluate what sort of training strategy works best for your
organization, but the takeaway here should be that you need to have an ongoing strat‐
egy—a strategy to address not only the initial training and expertise that your staff
will need, but also how you’re going to continue to reinforce past learning and intro‐
duce new learning.
Communication
Another area we often see overlooked is an intentional plan around communication.
As mentioned with training, communication is also not a “one-and-done” activity. It
is something that should not only be ongoing and consistent but also strategic and
exhaustive. In fact, we would argue that proper communication and an intentional
strategy around it are what fuels a powerfully effective data culture.
Top-down, bottom-up, and everything in between
When thinking about communication in terms of a data governance program, there
are multiple facets to consider. Two common facets of focus are top-down communi‐
cation of the governance program itself and its practices, standards, and expectations,
and bottom-up communication encompassing breaches and problems in governance
being bubbled up for resolution.
These two facets, while clearly important, are only pieces of the communication nec‐
essary to foster a company-wide data culture. These facets are focused simply on the
passage of governance information back and forth, not on how to develop and enrich
the culture of data privacy and security. For data culture, communication needs to
center around the tenets—what the culture for data is within the company, and how
each of those responsible for its success matter and are a part of it.
Additionally, while not specifically training per se, communication that bolsters a
company’s data culture can serve as a reminder and reinforcement not only of the
information covered in said trainings but also of the company’s overall vision and
commitment to its data culture.
Intention, Training, and Communications
|
199
Beyond Data Literacy
In the previous section we briefly discussed the value and impact of fostering caring
within the data culture. Here we will explore that more deeply and expand on why
this is such a fundamental part of a successful data culture.
Motivation and Its Cascading Effects
To be sure, education around what data is (namely, what different kinds of data are),
and how data should be treated is essential, but so too is the why. The question of why
data should be handled and treated with respect is an overlooked component of data
culture. Of course it is part and parcel for several of the user hats in the governance
space (legal and privacy tsars, to name a few), but it should be so for other hats as
well.
Motivation and adoption
This isn’t a text on psychology by any means; however, the power of motivation and
its impact on people’s likeliness to do this over that influences greatly whether a data
culture will be adopted or fall by the wayside.
For a data culture to be fully adopted, the motivation really needs to begin at the top
and permeate down from there. Take, for example, a company that needs to put into
place a governance program to abide by new data compliance and security laws. Sev‐
eral people (or teams) within the company know that their data collection and han‐
dling needs to meet higher standards to be in compliance and bring this need to the
C-level. For a governance program to be fully implemented (as we discussed earlier in
this chapter), the C-level folks need to buy into the idea that a governance program is
important and worth their support (including funding). Without C-level support for
a governance program, it’s highly likely that the program would fall short due to lack
of budgetary support and the absence of an ongoing advocacy of a data culture.
C-level involvement and belief in a governance program and implementation of a
data culture has many downstream effects.
First, there is the financial aspect. Without proper funding, a company does not have
the resources in terms of both headcount and tools but also in terms of education of
the data governance space and how to use said tools effectively. Not only does this
influence whether or not a governance program is executed effectively, but it also
affects how well people are able to do their jobs and how happy they are doing those
jobs. As we discussed in Chapter 3, people in the governance space wearing many dif‐
ferent hats and spreading their time across many different tasks—some of which
they’re ill-equipped to do—can lead to decreased job satisfaction and productivity
and ultimately high job turnover (which has its own implications in terms of sunk
cost and lost time).
200
| Chapter 9: Building a Culture of Data Privacy and Security
While C-level buy-in is clearly critical for its consequential effects, there is still the
need to cultivate motivation for best practices within the rest of the workforce. Here
again is where data culture becomes so important. First, the workforce needs to be
informed/educated on why proper data treatment and handling is not only necessary
but ethical. Of note here is that this is not where the reinforcement should stop. A
data culture continues to reinforce this value through the behavior of the company—
the training that is offered/required (and how often), the communications that are
sent out, the behavior of decision makers/those in positions of influence, and finally,
even the way the company is structured. A prime example of this would be a com‐
pany that touts a data culture of privacy and security and yet does not have dedicated
teams and/or resources that support that culture. If there is a disconnect between
what is internally marketed as the data culture and what actually exists to support that
culture, motivation and adoption of the data culture is not only less likely—it’s
improbable at best.
Maintaining Agility
Hopefully in the preceding sections we’ve driven home the need for a very thoughtful,
thorough, and structured approach to creating and cultivating a data culture. An
important aspect of this, one that should not be ignored in inception and creation of
a data culture, is how agility will be maintained. We have seen this be a highly prob‐
lematic area for many companies and would be remiss if we didn’t address its impor‐
tance, as it’s much easier to maintain agility than to create it after the fact.
Agility and Its Benefits on Ever-Changing Regulations
During the course of our research we encountered a very interesting use case relating
to the CCPA. While you may not be encountering compliance with this particular
regulation, the struggles one company is facing may give you some food for thought.
This company is a large retailer that has many different ingestion streams, as it sells
its products in many different locations across the United States. This company has
always struggled to keep track of not only the data that it collects itself, but also the
data it collects from third parties (e.g., a retailer sells its product directly to consumers
on its website but may also sell its product via other retailers, either in their stores or
via their websites).
Part of compliance with CCPA is to track down PII data for a California resident
regardless of where the item was purchased. For example, Ben lives in California but
purchased this company’s product at a large retailer while he was on vacation in Flor‐
ida. The transaction itself occurred in Florida, but because Ben is a California resi‐
dent, he can request that the product company find his purchase data and delete it.
Maintaining Agility
|
201
While many companies are now facing this conundrum, the company in this case
study has a highly structured and inflexible governance strategy—one that does not
easily lend itself to quickly adapting to these new requirements. As such, it recognizes
the need to revise its governance strategy, to allow the company to be more agile. The
main way it is spearheading this effort is to focus on building and nurturing a strong
data culture. With so many moving parts to its business (i.e., so much data coming
from so many different places), and with the likelihood that more regulations like
CCPA are on the horizon, this company feels that defining and building a culture that
will support (ever-changing) governance initiatives will be the key to its success. In
this way the company is creating (and has begun enacting) a comprehensive data
culture.
Requirements, Regulations, and Compliance
The most obvious reason for creating a culture around agility is the ever-changing
legal requirements and regulations that data must be in compliance with. To be sure,
the data regulations of today are likely to be less robust than the regulations of tomor‐
row. In Chapter 1 we discussed how the explosion of available data and its collection
has resulted in heaps and heaps of data just sitting there—much of it uncurated.
It is shortsighted and unfeasible to take an approach of simply “handling” the regula‐
tion(s) of the present. The ability to pivot once regulations change or pop up (because
they will) is not just a necessity—it can be made much easier if steps are taken at the
outset.
The Importance of Data Structure
A key aspect to consider when cultivating agility is to “set yourself up for success”—
again, this is a component that can be relied upon and enforced by the data culture.
Note that being “set up for success” is an intersection of how the data warehouse is
structured (the metadata that’s collected, the tags/labels/classification used, etc.) with
the culture that supports or enables the structure to run smoothly. These two in con‐
junction help to make it much easier to pivot when new regulations arise.
This is an area in which we’ve seen many companies struggle. Take, for example, data
that is stored based on the application from which it comes—say, sales data from a
specific company’s retail stores across the US. In the company’s current analytics
structure, this data is tagged in a certain way—perhaps with metadata relating to the
store, purchase amount, time of purchase, and so on.
Now let’s say that a new regulation comes along stating that any data from a customer
who lives in a certain state can be retained for only 15 days. How would this company
easily find all the sales data from customers who live in a certain state? Note that the
criteria is not sales with a certain state but customers, so a customer from state X
202
| Chapter 9: Building a Culture of Data Privacy and Security
could make a purchase in state Y, and if state X is part of the regulation, then that
customer’s data will need to be deleted after 15 days. The company in our example
will have a hard time complying quickly and easily with this regulation if it doesn’t
have the data structure set up to record the location of the purchaser (for the sake of
simplicity, we are assuming here that these are all credit card or mobile pay transac‐
tions that can be traced to each purchaser’s state of residence).
A process and culture from the outset to collect this kind of metadata (whether or not
it is labeled or tagged initially) will make it that much easier to find this class of data
and then tag/label it, and thus attach a retention policy.
Scaling the Governance Process Up and Down
While we have touted the importance of not only the right tools and process but also
the right players (hopefully provided by the buy-in from C-level), there undeniably
are times when things don’t go according to plan, and the process needs to scale (usu‐
ally down, although there is the use case that it could need to scale up).
It could be the case that a company loses headcount due to reorganization, restructur‐
ing, an acquisition, or even a change in data collection (in terms of actual data collec‐
ted and/or in the platform[s] used, tools for transformation, storage types and
locations, or analytics tools). Any of these will likely affect how a governance pro‐
gram functions, and that program needs to be elastic to accommodate for such
changes. The data culture—agreed upon, supported, and reinforced—will aid in suc‐
cessful elasticity.
One strategy we’d like to mention here is one we touched on in Chapter 2: making
sure that the most critical data is tended to first (again, this should be a part of the
data culture). It’s impossible to predict what requirements and regulations may come
in the future, but prioritizing data that is most critical to the business (as well as data
that has any relation back to a person or identifiable entity) will aid in your ability to
scale to accommodate whatever changes may come. At the absolute minimum there
should be a process in place (for example, all types of critical data are always tagged
or classified upon ingestion and/or are always stored in a certain location) that can
quickly address and tend to this category of data.
Interplay with Legal and Security
In this book we’ve discussed at length the different roles and/or “hats” involved in
data governance. When looking at the data culture specifically, two hats of note are
legal and security/privacy tsar, and the interplay between the two and their respective
teams is important.
Interplay with Legal and Security
|
203
Staying on Top of Regulations
Regardless of how a company is organized in terms of roles and who does what, there
must be someone who has the task of staying up to date on regulatory requirements
that the company may (or may not) need to be in compliance with. As we’ve previ‐
ously discussed, sometimes this is done by an internal legal representative (attorney)
and at other times this task falls on the privacy tsar or on someone else in security.
What’s important about this task is that it needs to be done early and often, not only
to ensure current compliance but also to help aid in the ability to ensure future com‐
pliance. As we’ve seen over the past 10 years, data handling standards and regulations
have greatly changed—the ability to be aware of what these changes might be and
how best to set up a data culture to flex as needed and be in compliance is critical.
As we’ve discussed several times throughout the book, and especially in Chapter 8,
having an auditing system in place will greatly aid you in monitoring your gover‐
nance strategy over time and also facilitate the task of being on top of compliance and
regulations, so that if (or when) you face an external audit, you will know that you are
in compliance and that nothing unexpected will be found.
Communication
Staying on top of regulations is, however, only half of the story. There must be a pro‐
cess in place for how changes in regulations will be discovered and how these will be
communicated to those who decide on how to proceed. In essence, there needs to be
constant communication back to the decision-making body about what regulations
might be coming up and/or whether there are any changes that need to be made to
current data handling practices in order to be in compliance.
It’s easy to see how this is a process that very much should be a part of the data cul‐
ture within a company.
Interplay in Action
An excellent recent example of this process in action is the reaction to GDPR. All
companies in the EU were well aware that this new regulation was coming and that
changes were needed in order to be compliant. The story was different for companies
in the US, however. During many of our interviews with US companies about their
data governance practices and their plans for the future in terms of changing regula‐
tions, we heard two different approaches: the first was to simply ignore new regula‐
tions until they become a requirement (so in the case of GDPR, not to address
compliance until it is a requirement for US companies); the second was to assume
that the most restrictive regulations in the world could become a requirement and
thus work toward being in compliance now, even before it’s required.
204
|
Chapter 9: Building a Culture of Data Privacy and Security
This takes a strong data culture that includes a robust interplay between the gathering
of legal requirements, a group deciding which requirements the company will comply
with, and another group actually carrying out the work of ensuring that compliance
is being met.
Agility Is Still Key
This point really hearkens back to the previous section on agility. It is likely that new
regulations will continue to emerge, and the flexibility of the data structure and sys‐
tem that a company builds would do well to have the potential to be easily modified
or to pivot when needed to accommodate such requirements.
Incident Handling
We’ve talked at length about the importance of process and communication in a suc‐
cessful data culture. One particular process we’d like to spend time unpacking is that
of incident handling—how are breaches in data governance handled, and who is held
responsible?
When “Everyone” Is Responsible, No One Is Responsible
During some of our initial research into how companies structure their data gover‐
nance strategy, one of the questions we asked was, “Who, at the end of the day, is
responsible for improper governance? Whose job is on the line?” Surprisingly, many
companies danced around this question and struggled to give us a direct answer of a
particular person and/or group who would be held accountable in the event of some‐
thing going wrong.
This might seem like an unimportant part of the data culture, but it is actually quite
critical. Earlier in this chapter we spoke about the importance of fostering an envi‐
ronment of caring and responsibility (and indeed this is important), but it also needs
to have a backbone—there must be a person(s) or group(s) who is culpable when
things go wrong. When this structure is lacking, governance becomes “everyone’s”
and yet “no one’s” responsibility. The culture needs to support everyone doing their
part in proper data handling and owning their portion of governance responsibilities,
and it also needs to identify who is the go-to—the end of the line, so to speak, for
specific aspects of the governance strategy.
For example, as part of the data culture, it is on every data analyst in a company to
know what is PII data and whether or not it should be accessed, and/or for what pur‐
poses it can be accessed. A good data culture will support these analysts with educa‐
tion and reinforcements for proper data handling. In the event that an analyst
mistakenly (or malevolently as the case may be) accesses and improperly uses PII
data, someone should be held accountable for that breach. It may be a data engineer
Incident Handling
|
205
in charge of enforcing access controls, or even a privacy tsar whose responsibility it is
to set up and manage access policies. In any event, there needs to be someone who
has, as part of their job, the responsibility for ensuring that things go right and
accepting the consequences when things go wrong.
We may be overstating this, but implementing and enforcing responsibility is key here.
People should also be trained in how to be responsible and what this looks like (and
what it doesn’t look like), and this should also be outlined in specific roles from the
outset. Literally listed within a role description’s key tasks should be what that role is
specifically responsible for in terms of data handling, and what the consequences are
for failure to carry out this task. Responsibility should be something that is defined
and agreed upon, as well as trained and communicated on, so that people are taught
how to be responsible and are also held responsible.
Importance of Transparency
Transparency is an oft-forgotten (or intentionally sidestepped) ingredient of gover‐
nance that warrants a deeper dive into not only why it’s important, but also why you
should keep it in mind and most certainly address it as part of building your data
culture.
What It Means to Be Transparent
To be sure, many companies and organizations do not want to disclose everything
about the ins and outs of their data governance structure, which is to be expected and
is not entirely problematic. There is value, however, in a certain amount of transpar‐
ency from an organization, in terms of what data it collects, how it uses the data (and
what for), and what steps/measures it takes to protect data and ensure proper data
handling.
Building Internal Trust
Earlier we touched on the importance of trust in building a data culture, and that
trust goes both ways—bottom up and top down. A key aspect of building that trust
and truly showing employees that it is part of the company’s data culture is to have full
transparency: transparency not only in the items related to data noted above (what
data is collected, how it’s used, governance processes, etc.) but also in what the
incident-handling strategy is. In the previous section we mentioned how important it
is to define the specific person or group that is responsible when things go wrong,
and just as there should be consequences for these folks, there should also be conse‐
quences (albeit different ones) for improper handling by anyone who touches data.
While it may make an organization feel uncomfortable to share so broadly about
when something has gone wrong and how it was handled internally, doing so builds
206
|
Chapter 9: Building a Culture of Data Privacy and Security
incredible trust within the company that what is being touted as the data culture is an
integral part of the entire company culture.
Another strategy that can help you build internal trust is that of enabling two-way
communication via a user forum, wherein users of data within your company are able
to voice their concerns and needs. We discussed earlier in this chapter the importance
of communication, but this is an additional facet that is not just informational (you
get to hear from the people actually using the data about why it might be wrong or
what could be better)—it also bolsters the data culture by making all in the organiza‐
tion feel that they’re being heard and are pieces of the greater whole that keeps the
governance program running smoothly.
Building External Trust
In creating a solid and successful data culture, focusing on the internal—the company
nuts and bolts—is obviously extremely important, but the data culture does not and
should not end there. The external perception and trust of a company/organization
should also be considered. Just as showing full transparency internally helps to build
and reinforce the data culture, what is communicated externally about data collec‐
tion, handling, and protection practices is also highly important.
In a sense, the customers of a company or organization are an additional extension of
its “culture.” Customers or consumers should also be considered when thinking about
building a data culture. It’s not just the actions and perceptions of your staff or
employees; it’s also the actions and perceptions of your consumers and/or customers,
and it’s their actions (generally driven by trust) that dictate whether or not they inter‐
act with your company or buy your product.
Providing full transparency externally regarding what data is collected, how it’s used,
how it’s protected, and what has been done to mitigate wrongdoing is critical in
building trust in a company/organization. There are, to be sure, cases in which cus‐
tomers/people have no choice but to choose one company or organization for a ser‐
vice/purchase/etc., but in the event that there is choice, customers/people are much
more likely to choose a company that they trust over one they don’t trust or are
unsure of.
The data culture in essence should not be just an internally decided-upon practice but
one that also includes how the company or organization sits in the world—what it
wants the world to know about how it handles data and its commitment to compli‐
ance and proper data handling.
Importance of Transparency
|
207
Setting an Example
It may sound lofty, but another aspect around the importance and maybe even the
power of transparency is that it can teach and/or inspire others to adopt similar data
governance practices and data culture.
If every company or organization were to create, implement, and enforce the gover‐
nance principles and practices we’ve laid out, not only would there be excellent gov‐
ernance and data culture within each organization, but there would also be a crosscompany, cross-organizational, cross-product, cross-geographical data culture.
Summary
Throughout the course of this text you have learned about all the aspects and facets to
be considered when creating your own successful data governance program. We hope
that we have educated you not only on all of the facets and features of data itself, gov‐
ernance tools, and the people and processes to consider, but also on the importance
of looking at your governance program in a much broader sense to include things
such as long-term monitoring and creating a data culture to ensure governance
success.
You should walk away from this book feeling that you know what makes up a gover‐
nance program (some aspects that you may have already heard of, and hopefully
some others that you hadn’t considered before) and that you know how to put those
pieces together to create and maintain a powerful, flexible, and enduring program
that not only meets but also exceeds regulation, compliance, and ethical and societal
standards.
208
|
Chapter 9: Building a Culture of Data Privacy and Security
APPENDIX A
Google’s Internal Data Governance
In order to understand data governance, it is good to look at a practical example of a
company with a deep investment in the topic. We (the authors) are all Google
employees, and we believe that in Google we have a great example of a deeply
ingrained process, and a fine example of tooling.
The Business Case for Google’s Data Governance
Google has user privacy at the top of its priorities and has published strong privacy
principles that guide us throughout all product development cycles. These privacy
principles include, as top priorities, respect for user’s privacy, being transparent about
data collection, and taking the utmost care with respecting the protection of users’
data, and they have ensured that good data governance is at the core of Google.
Before we dive into the specifics of data governance and management at Google, it is
crucial to understand the motivations behind Google’s data collection, and the use
case. This is a good approach for any undertaking. Google provides access to search
results and videos and presents advertisements alongside the search results. Google’s
income (while more diversified than a few years ago) is largely attributable to ads.
Given the importance of ads, a large portion of the effort in Google is focused on
making ads relevant. Google does this by collecting data about end users, indexing
this data, and personalizing the ads served for each user.
Google is transparent about this information: when someone uses Google services—
for example, when they do a search on Google, get directions on Maps, or watch a
video on YouTube—Google collects data to personalize these services. This can
include highlighting videos the user has watched in the past, showing ads that are
more relevant to the user depending on their location or the websites they frequent,
and updating the apps, browsers, and devices they use to access Google services. For
209
example, the results and corresponding ads might be different when a search is made
by a user on a mobile device and navigating in Google Maps versus a search made
from a desktop using a web browser. The personal information associated with the
user’s Google account is available if the user is signed in. This can include the user’s
name, birthday, gender, password, and phone number. Depending on the Google
property they are using, this can also include emails they have written and received (if
they are using Gmail); photos and videos they have saved (if they are using Google
Photos); documents, spreadsheets, and slides they have created (if they are using
Google Drive); comments they have made on YouTube; contacts they added in Goo‐
gle Contacts; and/or events on their Google Calendar. It is necessary that all this user
information is protected when using Google services.
Google, therefore, provides each user with transparency and control of how their per‐
sonal information is used. Users can find out how their ads are personalized by going
to Google Ad Settings and can control the personalization therein. They can also turn
off personalization and even take out their data. They can review their activity within
Google domains and delete or control activity collection. This level of transparency
and control is something that users expect in order to be comfortable with providing
a business with their personal information. Google maintains transparency around
what data it collects and how this information is used to generate the revenues men‐
tioned above. If you are collecting personal information and using it to personalize
your services, you should provide similar mechanisms for your customers to view,
control, and redact your use of their personal data.
Given all the information Google collects, it is not surprising that Google is publicly
committed to protecting that information and ensuring privacy. Google is meticulous
about external certifications and accreditations and provides tools for individual con‐
sumers to control the data collected about them.
The Scale of Google’s Data Governance
Google keeps some information about itself private—for example, how much data it
actually collects and manages. Some public information provides a general sense,
such as Google’s reported investment of $10 billion in offices and data centers in
2020.1 A third-party attempt at estimating Google’s data storage capacity using public
sources of information came up with 10EB (exabytes) of information.2
1 Carrie Mihalcik, “Google to Spend $10 Billion on Offices, Data Centers in US This Year”, CNET, February 26,
2020.
2 James Zetlen, “Google’s Datacenters on Punch Cards”, XKCD.
210
|
Appendix A: Google’s Internal Data Governance
Some further information about the Google data cataloging effort, its scale, and the
approaches taken to organizing data is described in the Google “Goods” paper.3 This
paper discusses the Google Dataset Search (GOODS) approach, which does not rely
on stakeholder support but works in the background to gather metadata and index
that metadata. The resulting catalog can be leveraged to further annotate technical
metadata with business information.
So, with this significant amount and variety of information, much of it likely sensi‐
tive, how does Google protect the data that it has collected and ensure that privacy
while keeping the data usable?
Google has published certain papers about the tools used, and we can discuss these.
Google’s Governance Process
There are various privacy commitments Google needs to respect and comply with, in
particular around user data: regulations, privacy policies, external communications,
and best practices. However, it is generally hard to:
• Make nontrivial global statements (e.g., all data is deleted on time)
• Answer specific questions (e.g., do you never record user location if setting X is
switched off?)
• Make informed decisions (e.g., is it OK to add this googler to that group?)
• Consistently assert rules or invariants at all times
The ideal state is one in which we have a comprehensive understanding of Google’s
data and production systems and automatically enforced data policies and obliga‐
tions. In the ideal state:
The lives of Google employees are easier
• Taking the privacy- and security-preserving path is easier than taking an insecure
path.
• Privacy bureaucracy is reduced via automation.
Google can do more…
• Developers use data without worrying about introducing security and privacy
risks.
• Developers can make privacy a product feature.
• Google can make data-supported external privacy statements with confidence.
3 Alon Halevy et al., “Goods: Organizing Google’s Dataset” (presented at SIGMOD/PODS’16: International
Conference on Management of Data, San Francisco, CA, June 2016).
Google’s Internal Data Governance
|
211
…while ensuring security and privacy
• Google can prevent privacy problems before they happen.
• Data obligations (policies, contracts, best practices) are objective and enforceable.
Every single feature Google produces and releases undergoes scrutiny by specialized
teams external to the core development team. These teams review the product from
the following aspects:
Privacy
• The team looks carefully at any user data collected and reviews justification for
the collection of this data. This includes reviewing what data, if collected, is going
to be visible to Google employees, and under what constraints. The team also
ensures that consent was provided for the data collected, and whether encryption
is applied and audit logging is enabled for that data.
• In addition, we look at compliance—we make sure that when users choose to
delete data, it is verifiably removed within Google’s committed SLAs, and that
Google has monitoring on all data retained (so that we can ensure minimal
retention is preserved).
• By verifying compliance and policies before even starting the collection of data,
we limit a significant number of challenges ahead of time.
Security
This is a technical review targeted at scrutinizing the design and architecture of
the code according to best practices to potentially head off future incidents
resulting from security flaws. Since most capabilities Google launches are
exposed to the web, and despite the fact that there are always multiple layers of
security, we always provide for an additional review. Web threats are present and
evolving, and a review by subject matter experts is beneficial to all.
Legal
This is a review from a corporate governance standpoint, making sure the prod‐
uct or service launched is compliant with export regulations and explicitly getting
a review from a regulation perspective.
(There are other approvers for a release, naturally, but we will focus on those related
to data governance.)
Google maintains additional certifications, with a common theme of most of those
being third-party verified.
212
|
Appendix A: Google’s Internal Data Governance
How Does Google Handle Data?
Much of the information Google holds goes into a central database, where it is held
under strict controls. We have already shared information about the likely contents of
this database; now let’s focus on the controls around this database.
Privacy Safe—ADH as a Case Study
ADH—or Ads Data Hub—is a tool Google provides that allows you to join data you
collect yourself (e.g., Google ad campaign events) with Google’s own data about the
same constituents. Yet it does so without breaching the privacy or trust of the individ‐
uals inspected. The ways ADH accomplishes this are indicative of the care Google
takes with respect to data. There are several mechanisms working in conjunction to
provide multiple layers of protection:
Static checks
ADH looks for obvious breaches, such as listing out user IDs, and blocks certain
analytics functions that can potentially expose user IDs or distill a single user’s
information.
Aggregations
ADH makes sure to respond only with aggregates, so that each row in the
response to a query corresponds to multiple users, beyond a minimal threshold.
This prevents the identification of any individual user. For most queries, you can
only receive reporting data on 50 or more users. Filtered rows are those omitted
from the results without notification.
Differential requirements
Differential requirements compare results from the current query operation
you’re running to your previous results, as well as rows from the same result set.
This is designed to help prevent the user from gathering information about indi‐
vidual users by comparing data from multiple sets of users that meet our aggre‐
gation requirements. Differential requirement violations can be triggered by
changes to your underlying data between two jobs.
Google’s Internal Data Governance
|
213
ADH Uses Differential Privacy
Businesses that advertise on Google often want to measure how well their marketing
is working. To do so, it is essential to be able to measure the performance of their ads.
A local restaurant that has advertised on Google will want to know how many people
who were served a restaurant ad actually came to the restaurant. How can Google
provide advertisers the ability to do customized analysis that aligns with the business
(e.g., how many patrons placed an online order after seeing the ad on Google), con‐
sidering such analysis will require joining information on who the ads are served to
with the restaurant’s own customer transactions database?
ADH uses differential checks to enable customized analysis while respecting user pri‐
vacy and upholding Google’s high standards of data security. Differential checks are
applied to ensure that users can’t be identified through the comparison of multiple
sufficiently aggregated results. When comparing a job’s results to previous results,
ADH needs to look for vulnerabilities on the level of individual users. Because of this,
even results from different campaigns, or results that report the same number of
users, might have to be filtered if they have a large number of overlapping users. On
the other hand, two aggregated result sets may have the same number of users—
appearing identical—but not share individual users and would therefore be privacysafe, in which case they wouldn’t be filtered. ADH uses data from historical results
when considering the vulnerability of a new result. This means that running the same
query over and over again creates more data for differential checks to use when con‐
sidering a new result’s vulnerability. Additionally, the underlying data can change,
leading to privacy check violations on queries thought to be stable.
These techniques are sometimes referred to as differential privacy.4
The ADH case study exemplifies Google’s cultural approach to handling data: beyond
the process part mentioned previously, Google has built a system that provides value
while at the same time ensuring safeguards and prevention techniques that put user
privacy at the forefront. Google has captured some of the capabilities in a “differential
privacy library”.
For another case study, consider the capabilities of Gmail. Google has built tools to
extract structured data from emails. These tools enable assistive experiences, such as
reminding the user when a bill payment is due or answering queries about the depar‐
ture time of a booked flight. They can also be combined with other information to do
seemingly magical things like proactively surfacing an emailed discount coupon
4 Some of the these techniques are described in a paper Google published entitled “Differentially Private SQL
with Bounded User Contribution”.
214
|
Appendix A: Google’s Internal Data Governance
while the user is at that store. All of the above is accomplished by scanning the user’s
personal email while still maintaining that user’s privacy. Remember, Google person‐
nel are not allowed to view any single email. This is presented in the paper “Anatomy
of a Privacy-Safe Large-Scale Information Extraction System over Email”. This capa‐
bility to scan information and make it accessible to the information’s owner, while at
the same time maintaining privacy, is accomplished through the fact that most emails
are business-to-consumer emails, and those emails from a single business to many
consumers share the same template. You can, without human intervention, backtrack
groups of emails to the business, generate a template that is devoid of any potential
information (differentiating between the boilerplate portions and the transient sec‐
tion), and then build an extraction template. The paper goes into more detail.
Google’s Internal Data Governance
|
215
APPENDIX B
Additional Resources
The following are some of the works we consulted while writing this book. This is not
intended to be a complete list of the resources we used, but we hope that some of you
will find this selection helpful as you learn more about data governance.
Chapter 4: Data Governance over a Data Life Cycke
• Association Analytics. “How to Develop a Data Governance Policy”. September
27, 2016.
• Australian Catholic University. “Data and Information Governance Policy”.
Revised January 1, 2018.
• Michener, William K. “Ten Simple Rules for Creating a Good Data Management
Plan”. PLOS Computational Biology 11, no. 10 (October 2015): e1004525.
• Mohan, Sanjeev. “Applying Effective Data Governance to Secure Your Data Lake”.
Gartner, Inc. April 17, 2018.
• Pratt, Mary K. “What Is a Data Governance Policy?”. TechTarget. Updated Febru‐
ary 2020.
• Profisee Group, Inc. “Data Governance—What, Why, How, Who & 15 Best Prac‐
tices”. April 12, 2019.
• Smartsheet, Inc. “How to Create a Data Governance Plan to Gain Control of
Your Data Assets”. Accessed February 26, 2021.
• TechTarget. “What Is Data Life Cycle Management (DLM)?”. Updated August
2010.
• USGS. “Data Management Plans”. Accessed February 26, 2021.
217
• Watts, Stephen. “Data Lifecycle Management (DLM) Explained”. The Business of
IT (blog). BMC. June 26, 2018.
• Wikipedia. “Data governance”. Last modified February 4, 2021.
• Wing, Jeannette M. “The Data Life Cycle”. Harvard Data Science Review 1, no. 1
(Summer 2019).
Chapter 8: Monitoring
• Alm, Jens, ed. Action for Good Governance in International Sport Organisations.
Copenhagen: Play the Game/Danish Institute for Sports Studies. March 2013.
• Cyborg Institute.“Infrastructure Monitoring for Everyone”. Accessed February
26, 2020.
• Ellingwood, Justin. “An Introduction to Metrics, Monitoring, and Alerting”. Dig‐
italOcean. December 5, 2017.
• Goldman, Todd. “LESSON—Data Quality Monitoring: The Basis for Ongoing
Information Quality Management”. Transforming Data with Intelligence. May 8,
2007.
• Grosvenor Performance Group. “How Is Your Program Going…Really? Perfor‐
mance Monitoring”. MAy 15, 2018.
• Henderson, Liz. “35 Metrics You Should Use to Monitor Data Governance”.
Datafloq. October 28, 2015.
• Karel, Rob. “Monitoring Data with Data Monitoring Tools | Informatica US”.
Informatica (blog). January 2, 2014.
• Pandora FMS. “The Importance of Having a Good Monitoring System? Offer the
Best Service for Your Clients” (blog post). September 19, 2017.
• Redscan. “Cyber Security Monitoring”. Accessed February 26, 2021.
• Wells, Charles. “Leveraging Monitoring Governance: How Service Providers Can
Boost Operational Efficiency and Scalability…”. CA Technologies. January 19,
2018.
• Wikipedia. “Data Lineage”. Last modified November 17, 2020.
218
| Appendix B: Additional Resources
Index
A
access control, 6-8
evolution of methods, 75
granularity, 137
identity and access management, 52
as policy, 41
regulation around fine-grained access con‐
trol, 26
access logs, 18
access management, 158-165
Access Transparency, 165
authentication, 158
authorization, 159
data loss prevention, 161
differential privacy, 164
encryption, 162
identity and access management, 52
policies, 160
user authorization and, 54
Access Transparency, 165
accountability, in Capital One data leak, 18-19
accounting (monitoring), 177
acquisitions, large companies and, 73
active monitoring systems, 189
addresses, deduplication of, 125
Ads Data Hub (ADH), 213-215
advertising, Google and, 209
agile data protection, 165-167
data lineage, 166
event threat detection, 167
security health analytics, 165
agility, data culture and
data structure and, 202
maintaining agility, 201-203
regulation and, 201
requirements, regulations, and compliance,
202, 205
scaling of governance, 203
AI (artificial intelligence)
case study: water meters, 118-120
data quality in, 117-120
alerting, 177
Amazon
bias in automated recruiting tool, 17
data-driven decision making, 14
Amazon Web Services (AWS) Nitro Enclaves,
149
analysis, real-time, 187
analytical data, life cycle of (see data life cycle)
analytics
business case for data culture, 195
in monitoring system, 188
ancillaries, 58
annotation of data, 123
API log, 136
approvers, defined, 7
artificial intelligence (AI)
case study: water meters, 118-120
data quality in, 117-120
attestations, in binary authorization, 157
attestors, in binary authorization, 157
audit logging, regulatory compliance and, 28
audit trail, 24
auditing (in monitoring), 177
auditing (of system)
data lineage and, 141
in framework, xv
authentication, 52, 158
219
Authoritative Vessel Identification Service
(AVIS), 19-22
authorization, 54, 159
automation
data protection, 96, 97
lineage collection, 135
AWS (Amazon Web Services) Nitro Enclaves,
149
B
Benford's Law, 144
best practices
data breach readiness, 171
data deletion process, 170
data protection, 167-173
electronic medical devices and OS software
upgrades, 170
physical security, 168
portable device encryption/policy, 170
separated network designs for data protec‐
tion, 168
big data analytics, 116
big data, defined, 29
binary authorization, 156
bottom-up communication, 199
business analyst, 64, 81
business case for data governance, xii, 194-196
analytics and bottom line, 195
building, 106
company persona/perception, 195
for Google's internal data governance, 209
importance of matching to data use, 121
top-down buy-in success story, 196
business value of data governance, 23-30
data governance considerations for organi‐
zations, 28-30
fostering innovation, 23
regulatory compliance, 26-28
risk management, 25
tension between data governance and
democratizing data analysis, 24
C
C-suite, 64
California Design Den, 15
Capital One data leak case study, 18-19
caring, as essential component of data culture,
197, 200-201
CCPA (California Consumer Privacy Act), 152
220
|
Index
effect on data governance, 9
use case, 201
central data dictionary, 66
classification (see data classification and orga‐
nization)
cloud
data governance advantages in public cloud,
30-34
data protection in, 146-149
moving data to, 30
multi-tenancy, 147
security surface, 147
virtual machine security, 148
virtual private cloud service controls, 155
Cloud Data Loss Prevention (DLP), 161
cloud-native companies, 67
CMEK (Customer Managed Encryption Keys),
163
Coast Guard, US, 19-22
Coast, Steve, 94
collection of data (see data collection)
company persona, 195
company, public perception of, 195
completeness of data, 129
compliance
audit logging, 28
business value of data governance, 26-28
challenges for companies, 76
data destruction and, 89
data governance considerations for organi‐
zations, 29
data lineage and, 141
data retention/deletion, 27, 97
gaming of metrics, 76
monitoring and, 177, 182
regulation around fine-grained access con‐
trol, 26
sensitive data classes, 28
Confidential Compute, 148
contact tracing, 60
Context-Aware Access, 161
costs
of bad data, 116
of cyberattacks, 185
of security breaches, 145
COVID-19
California's collection of data on, 116
location history privacy issues, 59-61
culture (see data culture)
Customer Managed Encryption Keys (CMEK),
163
customer support specialists, 64
customers, as extension of company culture,
207
customization, of monitoring system, 188
D
data
in business context, 75
decision-making and, 14-16
enhancing trust in, 5
evolution of access methods, 75
use case expansion, 14-16
data access management, xv
data accumulation, 29
data acquisition
defined, 87
identity and access management, 52
workflow management for, 52
data analysis
data quality in big data analytics, 116
tension between data governance and
democratizing data analysis, 24
data analyst, 63
data archiving
automated data protection plan, 97
as data life cycle phase, 89
in practice, 96
data assessment, 45
data breaches
Equifax, 167
healthcare industry, 171-173
physical data breach in Texas, 169
portable devices and, 170
readiness, 171
data capture, 87
data catalog, 24
data cataloging, xiv, 44
data change management, 141
data classes
in enterprise dictionary, 38-40
policies and, 40-43
data classification and organization, 43
access control and, 6-8
automation of, 43
data classification and organization, 43
data protection and, 146
in framework, xiv
data collection
advances in methods of, 10
increase in types of data collected, 13
in sports, 11-13
data completeness, 129
data corruption, 26
data creation
as data life cycle phase, 87
defining type of data, 95
in practice, 95
data culture
analytics and bottom line, 195
benefits of data governance to the business,
194-196
building a, 193-208
caring and, 200-201
communication planning, 199
data governance policy, 110
definition/importance of, 193
incident handling, 205
as intentional, 197
interplay with legal and security, 203
maintaining agility, 201-203
monitoring and, 191
motivation and adoption, 200
motivation and its cascading effects, 200
privacy/security and, 83, 193-208
staying on top of regulations, 204
training for, 197-199
transparency and, 206
data deduplication, 124
data deletion, 50-52
best practices, 170
regulatory compliance and, 27
data democratization, 24
data destruction
compliance policy/timeline for, 97
as data life cycle phase, 89
in practice, 97
data discovery and assessment, xiv
data enablement, 8
data encryption (see encryption)
data enrichment, 65
cloud dataset, 78
manual process of, 75
data entry, 87
data exfiltration, 153-158
secure code, 156
virtual private cloud service controls, 155
Index
|
221
zero-trust model, 157
data governance (generally)
in action, 17-22
advances in data collection methods, 10
basics, 1-35
business benefits of robust governance, xvii
business case for (see business case for data
governance)
business need for, xii
business value of, 23-30
classification/access control, 6-8
cloud-based data storage advantages, 30-34
data enablement/security versus, 8
defined, xiii
elements of, 2-8
enhancing trust in data, 5
ethical concerns around use of data, 16
growth of number of people working with
data, 10
growth of size of data, 9
holistic approach to, 4
improving data quality, 19-22
increased importance of, 9-17
internal data governance at Google, 209-215
managing discoverability/security/account‐
ability, 18-19
new laws/regulations around treatment of
data, 16
operationalizing (see operationalizing data
governance)
policy (see data governance policy)
process (see process of data governance)
program performance monitoring, 183-185
purpose of, 1
scaling of process, 203
tension between data governance and
democratizing data analysis, 24
tools (see tools of data governance)
data governance charter template, 101
data governance framework (see framework)
data governance policy
access management and, 160
changing regulatory environment, 109
complexity and cost, 109
considerations across a data life cycle,
108-111
data destruction and, 89
defined, 101
deployment time, 108
222
|
Index
developing a policy, 103
importance of, 102
location of data, 110
organizational culture, 110
roles and responsibilities, 105
step-by-step guidance, 106-108
structure of, 103-105
data handling, at Google, 213-215
data in flight, governance of, 133-142
data transformations, 133
lineage, 134-140
policy management, simulation, monitor‐
ing, change management, 141
security, 151
data infrastructure, complexity of, 30
data life cycle, 85-111
applying governance over, 92-100
case study: information governance, 99
data archiving phase, 89, 96
data creation phase, 87, 95
data destruction phase, 89, 97
data governance frameworks, 92-94
data governance in practice, 94-97
data governance policy considerations,
108-111
data governance throughout, 85-111
data management plan, 90-92
data processing phase, 88, 95
data storage phase, 88, 96
data usage phase, 88, 96
defined, 85
example of how data moves through a data
platform, 97
management of, 90-92
operationalizing data governance, 100
phases of, 86-90
data lifecycle management (DLM), 90-92
data lineage, 134-140
agile data protection and, 166
auditing and compliance, 141
collecting, 135
data protection planning and, 144
defined, 46
key governance applications that rely on lin‐
eage, 139
monitoring, 180
policy management, simulation, monitor‐
ing, change management, 141
as temporal state, 138
tracking of, 46, 128
types of, 136
usefulness of, 135
data locality, 31
data loss prevention, access management and,
161
data maintenance (see data processing)
data management plan (DMP), 90-92
data outliers (see outliers)
data owner, 62, 80
data policies
data classes and, 40-43
per-use case, 42
data processing
as data life cycle phase, 88
documenting data quality expectations, 95
in practice, 95
data profiling, 45
data completeness, 129
data deduplication, 124
data quality and, 124-129
dataset source quality ranking for conflict
resolution, 129
lineage tracking, 128
merging datasets, 129
outliers, 127
data protection, 143-174
as framework element, xv
automated, 96
automated data protection plan, 97
best practices, 167-173
classification and, 146
in the cloud, 146-149
data breach readiness, 171
data deletion process, 170
data exfiltration, 153-158
identity/access management, 158-165
keeping data protection agile, 165-167
level of protection, 145
lineage and quality, 144
network security, 151
physical security, 149-152
planning, 143-146
portable device encryption/policy, 170
security in transit, 151
data provenance (see data lineage)
data quality, 113-131
AI/ML models and, 117-120
annotation, 123
big data analytics and, 116
cascading problems with tribal knowledge,
124
data protection planning and, 144
data quality management as part of frame‐
work, xv
defined, 113
documenting data quality expectations, 95
importance of, 114-120
improving, 19-22
management processes in enterprise dictio‐
nary, 46
matching business case to data use, 121
monitoring, 179
prioritization, 123
profiling, 124-129
reasons for including in data governance
program, 121
scorecard for data sources, 123
techniques for, 121-130
unexpected sources for, 129
at US Coast Guard, 19-22
data recovery, automated, 96
data retention, 50-52
case study, 51
regulatory compliance and, 27
data scientist, 63
data security, 8
Capital One data leak case study, 18-19
creating culture of, 83
data governance versus, 8
healthcare industry security breaches,
171-173
key management and encryption, 47-49
physical, 149-152
security monitoring, 185
Zoom bombing, 152
data segregation
ownership by line of business, 80-82
within storage systems, 77-79
data sensitivity, 135
data steward, 7, 63, 81
data storage
automated data protection and recovery, 96
as data life cycle phase, 88
in practice, 96
key management and encryption, 47-49
data structure, agility and, 202
data theft, managing risk of, 25
Index
|
223
data transformations, 133
data usage
access management and, 96
as data life cycle phase, 88
in practice, 96
data validation, 133
data-driven decision making, 14
datasets
"views" of, 82
merging, 129
source quality ranking for conflict resolu‐
tion, 129
debugging, 139
decision-making, data and, 14-16
deduplication, 124
deletion of data (see data deletion)
democratization of data, 24
differential privacy, 59, 164, 214
digital-only (cloud-native) companies, 67
director of data governance (see privacy tsar)
discoverability, 18-19
DLM (data life cycle management), 90-92
DLP (Cloud Data Loss Prevention), 161
DMP (data management plan), 90-92
documenting data quality expectations, 95
E
education, when operationalizing data gover‐
nance, 108
electronic medical devices, 170
employees (see people, data governance and)
encryption, 151
access management and, 162
key management and, 47-49
portable devices and, 170
enterprise dictionary, 37
data assessment/profiling, 45
data cataloging/metadata management, 44
data classes and policies, 40-43
data classes in, 38-40
data classification and organization, 43
data quality management processes, 46
data retention/deletion, 50-52
key management and encryption, 47-49
lineage tracking, 46
workflow management for data acquisition,
52
Equifax data breach, 167, 185
ethics, data use and, 16
224
|
Index
ETL (extract-transform-load), 116, 133
European Union (EU), 3
AI regulations, 17
Spotify regulations, 3
event threat detection, 167
exfiltration (see data exfiltration)
Experian, 116
external auditor, 64
external trust, 207
extract-transform-load (ETL), 116, 133
extraction, defined, 133
F
filters, 121
fine-grained access control, 26
framework, xiv
accountability, 107
and data life cycle, 92-94
G
GDPR (General Data Protection Regulation),
xi, 26
data retention rules, 51
data sovereignty measures, 31
effect on data governance, 9
right to be forgotten, 76
US companies and, 204
Gmail, 214
Google
Ads Data Hub case study, 213-215
business case for internal data governance,
209
data handling at, 32, 213-215
governance process, 211-212
internal data governance at, 209-215
protocol buffers, 151
scale of internal data governance, 210
Google Cloud
encrypted data in, 48
Shielded VM, 148
Google Dataset Search (GOODS), 211
governance manager (see privacy tsar)
governance process (see process of data gover‐
nance)
governors, defined, 7
graphic visualization, in monitoring system,
188
gRPC framework, 151
guiding principles, developing/documenting,
106
H
"hats"
and company structure, 74
Health Insurance Portability and Accountabil‐
ity Act of 1996 (HIPAA), 16
healthcare industry
cascading data quality problems with tribal
knowledge, 124
data protection best practices, 167
need for comprehensive data governance,
196
security breaches, 171-173
Herzberg, Elaine, 16
highly regulated companies, 69-71
I
identity and access management (IAM) sys‐
tems, 52, 147
identity-aware proxy (IAP), 160
IMPACT (teacher ranking system), 76
incident handling, 205
infotypes, 37
inheritance, 141
innovation, fostering with data governance, 23
internal data governance, at Google, 209-215
internal trust, building, 206
J
Jacob, Oren, 50
K
k-anonymity, 164
key management, 47-49
key management system (KMS), 163
knowledge workers, 8
L
labeling of resources in public cloud, 33
laptops, 170
large companies, 72
legacy companies, 66
legal "hat", 58
legal issues (see regulations; compliance)
life cycle (see data life cycle)
lineage (see data lineage)
lineage graph, 136
lineage tracking, 128
in enterprise dictionary, 46
time/cost of, 47
location of data, 110
M
machine learning (ML)
data governance and, 4
data quality in, 117-120
management buy-in, 107
maritime domain awareness (MDA), 21
Mars Climate Orbiter, 100
MaxMind, 115
medical records, 167
mergers and acquisitions, large companies and,
73
metadata
analytics in legacy companies and, 66
data enrichment and, 65
data lineage and, 135
metadata catalog, 24
metadata management
in enterprise dictionary, 44
as part of framework, xiv
metrics, gaming of, 76
Microsoft Azure, 148
misuse of data, 25
ML (machine learning)
data governance and, 4
data quality in, 117-120
monitoring, 175-191
compliance monitoring, 182
criteria for, 189
data lineage monitoring, 180
data quality monitoring, 179
defined, 175
important reminders for, 190
key areas to monitor, 179-187
program performance monitoring, 183-185
reasons to perform, 176-178
security monitoring, 185
system features, 187-189
monitoring system
analysis in real-time, 187
customization in, 188
features of, 187-189
graphic visualization in, 188
notifications in, 187
Index
|
225
reporting/analytics in, 188
system alerts, 187
multi-tenancy, 147, 153
N
names, deduplication of, 125
NASA, 100
National Football League (NFL), 11-13
National Health Service (NHS), 167
natural experiments, 99
network security, 151
Newett, Edward, 3
Next Gen Stats (NGS), 11
NFL (National Football League), 11-13
NHS (National Health Service), 167
Nike Zoom Vaporfly 4% ad campaign, 99
Nitro Enclaves, 149
normalization, 45, 134
notifications, in monitoring system, 187
O
ontologies, 107
OpenStreetMap (OSM), 94
operating model, 107
operationalizing data governance, xvi, 100
considerations across a data life cycle,
108-111
data governance policy defined, 101
developing a policy, 103
importance of a data governance policy, 102
policy structure, 103-105
roles and responsibilities, 105
step-by-step guidance, 106-108
OS software upgrades, 170
OSM (OpenStreetMap), 94
outliers, 45, 114, 127
P
passive monitoring systems, 189
people, data governance and, 57-65
considerations and issues, 74-77
creation of "views" of datasets, 82
culture of privacy/security, 83
data definition, 75
data enrichment, 65
data segregation and ownership by line of
business, 80-82
226
|
Index
data segregation within storage systems,
77-79
for physical security, 149
growth of number of people working with
data, 10
"hats," defined, 58
"hats" versus "roles" and company structure,
74
legal "hat", 58
people–data governance process synergy,
73-84
regulation compliance, 76
roles, responsibilities, and "hats", 57
tribal knowledge and subject matter exerts,
74
per-use case data policies, 42
perimeter network security model, 151
personally identifiable information (PII) (see
PII)
personnel (see people, data governance and)
physical security, 149-152
best practices, 168
network security, 151
physical data breach in Texas, 169
security in transit, 151
PII (personally identifiable information)
data class hierarchy and, 39
data deletion and, 50
defined, 40
lineage collection, 136, 139
policy book (see enterprise dictionary)
policy(ies) (see data governance policy)
portable devices, 170
prioritization of data, 123
privacy
creating culture of, 83
differential privacy, 164
privacy tsar, 59
community mobility reports, 59
exposure notifications, 60
process of data governance, 65-73
cloud-native companies and, 67-69
at Google, 211-212
highly regulated companies, 69-71
large companies, 72
legacy companies and, 66
legacy retail companies, 68
people and (see people, data governance
and)
scaling of, 203
small companies, 71
profiling (see data profiling)
program performance monitoring, 183-185
ProPublica, 17
protocol buffers, 151
provenance (see data lineage)
public cloud
Capital One data leak, 18
data governance in, 30-34
data locality and, 31
ephemeral computing, 32
labeled resources, 33
reduced surface area, 32
security issues, 34
as serverless and powerful, 32
public perception of company, 195, 207
purging (see data destruction)
Q
quality (see data quality)
quality ranking, 129
R
RACI matrix (responsible, accountable, consul‐
ted, and informed), 105
regulations
agility and compliance, 201, 205
around treatment of data, 16
communicating regulatory changes, 204
compliance (see compliance)
data destruction and, 89
data governance considerations for organi‐
zations, 29
data governance policy regulatory environ‐
ment, 109
highly regulated companies, 69-71
staying on top of, 204
regulatory compliance (see compliance)
reporting, in monitoring system, 188
responsibility, incident handling and, 205
retail companies
compliance use case, 201
data governance process for, 67-69
legacy companies, 68
regulation issues, 202
retention of data (see data retention)
right to be forgotten, 76
risk management, 25
S
Safari Books Online, 14-15
scaling of governance process, 203
Science Applications International Corp (SAIC)
data breach, 169
scorecard, for data sources, 123
security (see data protection; data security)
security breaches (see data breaches)
security monitoring, 185
security teams, 147
self-driving cars, 16
sensitive data classes, 28
sensitivity of source data, 135
Shielded VM, 148
signers, in binary authorization, 157
small companies, data governance in, 71
SMEs (subject matter exerts), 74, 81
Society of Vertebrate Paleontology, 121
software upgrades, 170
sports, advanced data collection in, 11-13
Spotify, 2-4
storage systems, data segregation within, 77-79
Strava, 99
streaming data (see data in flight, governance
of)
street addresses, deduplication of, 125
subject matter exerts (SMEs), 74, 81
Susman, Gayln, 50
system access logs, 18
system alerts, 187
T
tagging, 33
taxonomies, developing, 107
teacher rating systems, 76
technology stack, 107
Telco, 25
tenets, of company, 197
test dataset, 117
text data filters, 121
theft (see data theft)
tools of data governance, 37-55
data assessment/profiling, 45
data cataloging/metadata management, 44
data classes and policies, 40-43
data classes in, 38-40
data classification and organization, 43
data quality management processes, 46
data retention/deletion, 50-52
Index
|
227
enterprise dictionary, 37
identity and access management, 52
key management and encryption, 47-49
lineage tracking, 46
user authorization and access management,
54
workflow management for data acquisition,
52
top-down communication, 199
Toy Story 2 (movie), 50
training, 108
training dataset, 117
transactional systems, 86
transformations (see data transformations)
transparency
building external trust, 207
building internal trust, 206
data culture and, 206
meaning of, 206
setting an example, 208
tribal knowledge, 74, 124
TRICARE data breach, 169
trust
and purpose of data governance, 1
enhancing trust in data, 5
external, 207
internal, 206
U
use cases, for data
228
|
Index
expansion in, 14-16
importance of use case and policy manage‐
ment, 43
per-use case data policies, 42
user authentication, 52, 158
user forums, 207
user roles, company structure and, 74
users, defined, 7
V
validation dataset, 117
validation of data, 133
views of datasets, 82
virtual machine (VM), 148
virtual private cloud service controls (VPCSC), 155
W
Washington, DC, 76
water meter case study, 118-120
workflow management, for data acquisition, 52
workforce (see people, data governance and)
Z
zero-trust model, 157
Zippilli, Dom, 19
Zoom bombing, 152
About the Authors
Evren Eryurek, PhD, is the leader of the data analytics and data management portfo‐
lio of Google Cloud, covering Streaming Analytics, Dataflow, Beam, Messaging
(Pub/Sub & Confluent Kafka), Data Governance, Data Catalog & Discovery, and Data
Marketplace as the director of product management.
He joined Google Cloud as the technical director in the CTO Office of Google Cloud,
leading Google Cloud in its efforts toward Industrial Enterprise Solutions. Google
Cloud business established the CTO Office and is still building a team of the world’s
foremost experts on cloud computing, analytics, AI, and machine learning to work
with global companies as trusted advisors and partners. Evren joined Google as the
first external member to take a leadership role as a technical director within the CTO
Office of Google Cloud.
Prior to joining Google, he was the SVP & software chief technology officer for GE
Healthcare, a nearly $20 billion segment of GE. GE Healthcare is a global leader in
delivering clinical, business, and operational solutions, with its medical equipment,
information technologies, and life science and service technologies covering settings
from physician offices to integrated delivery networks.
Evren began his GE career at GE Transportation, where he served as general manager
of the software and solutions business. Evren was with Emerson Process Management
group for over 11 years, where he held several leadership positions and was responsi‐
ble for developing new software-based growth technologies for process control sys‐
tems and field devices, and coordinating cross-divisional product execution and
implementation.
A graduate of the University of Tennessee, Evren holds master’s and doctorate
degrees in nuclear engineering. Evren holds over 60 US patents.
Uri Gilad is leading data governance efforts for the data analytics within Google
Cloud. As part of his role, Uri is spearheading a cross-functional effort to create the
relevant controls, management tools, and policy workflows that enable a GCP cus‐
tomer to apply data governance policies in a unified fashion wherever their data may
be in their GCP deployment.
Prior to Google, Uri served as an executive in multiple data security companies, most
recently as the VP of product in MobileIron, a public zero trust/endpoint security
platform. Uri was an early employee and a manager in CheckPoint and Forescout,
two well-known security brands. Uri holds an MS from Tel Aviv University and a BS
from the Technion—Israel Institute of Technology.
Valliappa (Lak) Lakshmanan is a tech lead for Big Data and Machine Learning Pro‐
fessional Services on Google Cloud Platform. His mission is to democratize machine
learning so that it can be done by anyone anywhere using Google’s amazing infra‐
structure (i.e., without deep knowledge of statistics or programming or ownership of
lots of hardware).
Anita Kibunguchy-Grant is a product marketing manager for Google Cloud, specifi‐
cally focusing on BigQuery, Google’s data warehousing solution. She also led thoughtleadership marketing content for Data Security & Governance at Google Cloud.
Before Google, she worked for VMware, where she managed awareness and go-to
market programs for VMware’s core Hyper-Converged Infrastructure (HCI) product,
vSAN.
She has an MBA from MIT Sloan School of Management and is passionate about
helping customers use technology to transform their businesses.
Jessi Ashdown is a user experience researcher for Google Cloud specifically focused
on data governance. She conducts user studies with Google Cloud customers from all
over the world and uses the findings and feedback from these studies to help inform
and shape Google’s data governance products to best serve the users’ needs.
Prior to joining Google, Jessi led the enterprise user experience research team at TMobile, which was focused on bringing best-in-class user experiences to T-Mobile
retail and customer care employees.
A graduate of both the University of Washington and Iowa State University, Jessi
holds a bachelor’s in psychology and a master’s in human computer interaction.
Colophon
The animal on the cover of Data Governance: The Definitive Guide is the Pakistan
tawny owl (Strix aluco biddulphi). While tawny owls are common throughout Europe,
Asia, and northern Africa, this subspecies is specifically found in the northern parts
of Afghanistan, Pakistan, and India, as well as in Tajikistan and Kyrgyzstan. These
owls prefer temperate deciduous forests, or mixed forests with some clearings. They
may also roost in shrub land, orchards, pastures, or urban parks with large trees—
anywhere with enough foliage to keep them well hidden during the day.
Tawny owls tend to have medium-sized brown or brown-gray bodies with large
round heads and deep black eyes. Pakistan tawny owls have a distinctive gray color‐
ing with a whitish base and strong herringbone pattern below the head and mantle.
They are also believed to be the largest subspecies, with wingspans around 11 to 13
inches; females are often slightly larger than their male counterparts. These owls are
strictly nocturnal and are not often seen in daylight. They are carnivorous and hunt
for small mammals, rodents, reptiles, birds, insects, and fish from dusk to dawn. They
do not migrate and are considered mature at the end of their first year.
Tawny owls are monogamous and pair for life. Breeding season is from February to
July, and they nest in tree holes, among rocks, or in the crevices of old buildings. The
female incubates two to four eggs for a month. New hatchlings are altricial and can‐
not fend for themselves or leave the nest for the first 35 to 40 days; they are fed by the
mother with food brought by the father. Juveniles stay with their parents for about
three months after fledging, at which point they may disperse after breeding to estab‐
lish new territory within their local range. Pakistan tawny owls can be highly territo‐
rial and defend an area of about one thousand square meters year-round.
Tawny owls are excellent hunters, and while their eyesight may not be much better
than a human’s, they have excellent directional hearing and can swivel their heads
almost 360 degrees when tracking prey. They typically live 4 years in the wild—the
oldest wild tawny owl ever recorded lived more than 21 years! Tawny owls are consid‐
ered a species of least concern by the IUCN. Many of the animals on O’Reilly covers
are endangered; all of them are important to the world.
The cover illustration is by Karen Montgomery, based on a black and white engraving
from Meyers Kleines Lexicon. The cover fonts are Gilroy Semibold and Guardian
Sans. The text font is Adobe Minion Pro; the heading font is Adobe Myriad Con‐
densed; and the code font is Dalton Maag’s Ubuntu Mono.
There’s much more
where this came from.
Experience books, videos, live online
training courses, and more from O’Reilly
and our 200+ partners—all in one place.
©2019 O’Reilly Media, Inc. O’Reilly is a registered trademark of O’Reilly Media, Inc. | 175
Learn more at oreilly.com/online-learning
Download